[Feature Request] Write a trajectory from an AtomGroup and an Array

[jdetle]

@kain88-de suggested I open an issue for discussion about this in a #896 PCA comment. It would be convenient to have a function to write an atom group to its own trajectory given an array containing its updated positions would be nice.

Here we can discuss where such a function should exist and how we should go about implementing it.

[orbeckst]

Can you add a link to the discussion where this came up?

(Note: If you add a link to the "commented X hours ago" or similar time information, it formats nicely, like so #915 (comment) even though you pasted https://github.com/MDAnalysis/mdanalysis/issues/915#issue-167486514)

[kain88-de]

See #896 (comment)

For the PCA it came up to make an inverse projection from the PCA-space back to the conformation space. Instead of just giving an array of numbers I thought it would be nice to have an option to write out a trajectory for a given numpy array. A colleague is calculating new points in PCA-space and converting them back into conformation space, it would be convenient for him to directly convert that to a trajectory file on disk.

I thought it might be useful in other circumstances as well to convert a numpy array to a trajectory. If others think that is useful as well we can straight go to implement it in a more general way then just for the PCA class. I'm also not sure where such a function would best live so that others can find it.

[jdetle]

What about having an analysis.tools module? Stuff that could go here would be:

• atomgroup_write
• all_frames analysis
• custom metrics (maybe we could move the qcp.RMSD here?)

[richardjgowers]

So is this just essentially...

traj = np.array()  # shape (nframes, natoms, 3)
ag = u.select_atoms()  # size natoms

with mda.Writer('out.xtc', natoms=len(ag)) as w:
    for frame in traj:
        ag.positions = frame
        w.write(ag)

Or is it more complicated than this?

[kain88-de]

Nope that is it. It should have some guards for the shape of traj as well but not more.

[jbarnoud]

Couldn't the memory reader introduced in #797 do the trick?

[jdetle]

Reading the code it looks promising. Does an atomgroup satisfy all the requirements for a topology? (Sorry I don't really fully comprehend readers/universes yet...)

u = mda.Universe(atomgroup, np_array_of_coordinates, inmemory=True)

looks like it should work, no?

[richardjgowers]

@jdetle I think an AtomGroup should be enough.. I hadn't thought of that before. Would need a parser but that should be pretty simple.

[kain88-de]

Couldn't the memory reader introduced in #797 do the trick?

If we only use it inside of MDAnalysis that would be enough. But if I want to use the trajectory with other tools I still need to save it. I would find it useful for myself and can write a function for my own use. But we can just skip the saving part.

Creating an inmemory reader from an Array though should definitely be on the list.

[jbarnoud]

The memory reader from #797 being a reader, you should be able to pass it to a writer. Shouldn't you?