What is the best way to compare 3Dflex across datasets, such as ligand-bound vs unbound, and ensure that the observed changes in dynamics are due to the ligand rather than to differences in upstream processing (making of mesh, masks, etc.)?
1 Like
Hi Ricardo,
indeed, it is very important to compare comparable datasets! Are the particle stacks similar in terms of latent space distribution? Is the resolution similar? Is the number of particle similar too?
I used 3DVA to compare ligand-bound vs apo datasets which gave me good results (ensuring the above). I used phenix.varref to post-process the maps, refining an ensemble of PDBs in you ensemble of maps, and allowing you to visualize what is really going on (rotation, translation etc…). This way, I could see that some movements were conserved between the 2 datasets, and some were affected in amplitude giving me hints on the ligands’ impact on the conformational spectrum of the protein.
1 very important point: you are not looking at a change in dynamics! You are looking at a change in the conformational spectrum of your protein, meaning the different poses that can be extracted from your particle stack. To be able to use the word dynamics, you need MD and add time to these measurement, and if you can add force to it the you may use the word mechanics. Even though it looks like a movement, and in a way it is very close to it as the particles are close to each other, we don’t know the reaction coordinate linking the particles together but the “movement” is computed. So our job is to integrate thes results with others to reconstitute a plausible movement, but conditional verbing should be used to describe what is going on.
Well, I guess yes, it is important to keep all the parameters identical if you plan to compare stuff. The tricky part of this kind of analysis is that it does not necessarily give the same importance to the same kind of heterogeneity across different runs. Protein movements are often very complex and too difficult to describe in a single component. I usually look at the spread of the latent dimensions plot (the blue shotgun mark) for that kind of comparisons. Then focus on a single kind of movement and try to quantify it somehow (measure the amplitude of twists or shifts). But then across samples I check at least the 3 first modes to see how they are interpreting stuff. 3DVA is usually easier to interpret IMO, based on that marvelous colored interactive plot. And the resulting maps are buildable if you are in Clusters mode - contrary to 3DFlex which produces deformations of the initial map. By the other hand, because 3DVA is making clusters, it might underestimate the maximal amplitudes, I suppose…
I’d really like to read what the CryoSparc crew suggests on this topic.
Cheers