Details about pose / rotation format

[ahundt]

@mpharrigan, thanks for writing your Tensorflow RMSD article article. This looks very cool & I'm excited to learn about the RMSD distance metric!

Based on my reading I have a couple of questions, because I'm unsure of the data type/ordering you used. Does the following sound correct?

Some baseline info for what I'm familiar with:

• Positions are, of course, xyz format with 3 coefficients
• quaternions are often represented in wxyz and xyzw format

You describe the trajectories as follows:

# traj = np.array(...) [shape (n_frames, n_atoms, 3)]
traj -= np.mean(traj, axis=1, keepdims=True)

1. Does mean the trajectories are simply a series of dimension [n_frames, n_atoms, 3] where the 3 is xyz coordinates for each atom at each time step?
2. In what format are rotations represented in the data or where are they defined?
3. If I have two arrays, predicted and ground truth with format xyz_qwxyz and shape [n_pred, 7], would RMSD be a good way to measure the loss between them?
4. Do you think it is feasible to adapt the code for applying this distance metric on data in xyz_qwxyz format with shape [n_pred, 7]?

def key_matrix(r):

5. what format is r in key_matrix(r), is it just a vector xyz relative to 0?

I may be misunderstanding how the mathematical abstraction of the molecules works, because I don't typically work with molecular dynamics. In my case, I happen to have a robot with 6 rotating joints, but they basically work like a really, really big molecule so I thought this might be an interesting algorithm to take a look at! :-)

Thanks again