■

ここ数日、妙に多忙な上、あちらこちらから変な解析をしているメールが来て、間違っているものを放置しておくわけにもいかず、振り回された感があるが、それなりに興味深い結晶学的事例ばかりで、終わってみれば面白かった。

Do not refine against maps from MultiBody (relion_flex_analyze) or 3DVA

Replying to this tweet about varreef.

I'm not sure if outputs from 3DVA or Mulltibody are suitable for atomic model refinement. These maps are generated with very crude approximations and show many artifacts. There are no half map pairs for validation.

Multibody maps are generated by combining regional maps (called "bodies") moving as rigid bodies. Thus, interface regions are not well defined. Different bodies might have different resolutions and SNRs. Why don't you refine atomic models against individual bodies and show the distribution of rigid body parameters in run_data.star?

3DVA is a voxel-wise linear model, which cannot express large motion. In the movie above, we can see the central helix (which has N581 at the end) becomes thinner and fatter. Can we really refine against such maps and quantitatively say the movement is 6.2 Å?

One can select particles based on their position on a histogram (see Fig. 2E/F in the eLife paper). I guess 3DVA has a similar feature. Making 3D reconstructions from such subsets of particles and refining atomic models against them are a better strategy.

I am not talking about resolutions. My concern is inherent limitations of the output maps from 3DVA or MultiBody (relion_flex_analyze). The approximations (extrapolations) used in these tools break down at large motions. These maps can be wrong even at low resolutions. Clearly maps with crashing domains cannot be physical and one should not refine models (even secondary structures) against them. Maps from relion_flex_analyze are not "real" maps. They are just superpositions of each body at extrapolated locations/poses. Isolate particles and run Refine3D! Also note that histograms are projections onto one PC axis. You might have large conformational flexibilities along other dimensions. This will lead to very blurry maps if you select particles by looking at only one PC axis.

My bottom line is that your tool should be applied to subset reconstructions, not to interpolated/extrapolated maps from 3DVA or MultiBody.