Accuracy of the Atoms to Voxel Model

mjo22 / cryojax

Cryo electron microscopy image simulation and analysis built on JAX.

https://mjo22.github.io/cryojax/

GNU Lesser General Public License v2.1

30 stars 13 forks source link

Accuracy of the Atoms to Voxel Model #177

Open ehthiede opened 8 months ago

ehthiede commented 8 months ago

The atoms to voxel model is largely unproven, and we likely need some work to make it more physical. A nonexhaustive list of things to consider:

[x] We need to resolve the prefactors mention in #167.
[ ] We are missing hydrogens in our PDB loader (!)
[ ] We don't have support for charged atoms.
[ ] We don't have solvent effects included in our model.
[x] The role of the $B$ in this reference should be studied more fully, or at least supported.

mjo22 commented 8 months ago

As a part of this, it would be good to re-vamp something like the AtomCloud. If we eventually want to optimize over parameters in an atomic representation, we should be storing these parameters in the leaves of some pytree (here, an eqx.Module).

ehthiede commented 8 months ago

The question of "what do we do with AtomCloud" is a good one... I think you might be suggesting merging it with an equinox module containing the data of the atomic representation. I think this is appealing down the road, but I kind of want to make those API changes a separate issue for future reference.

mjo22 commented 8 months ago

That makes sense!

mjo22 commented 6 months ago

I believe we now have resolved item 1 on this list with closing of #167 and merging of #218. Have written up the math in #216. I will add to this list

[x] Support experimental b factors when building a potential. Maybe they are already supported, but I am not sure if the gaussians used for the scattering factors and the gaussians for PDB b factors have the same convention.