Dependency issues

[carbondrop-nick]

Super cool, but hard to get going. On top of the stated dependencies I also had to add a lot of modules to my environment to get this working:
ipdb einops ml-collections dm-tree ipywidgets jupyter einx torch_geometric tmtools openmm POT rdkit mdtraj pdbfixer

Seems to be missing "sampling" module. The imported configs, inference, and data.loader point to modules that don't exist (unless you meant the ones under Pretrain?)

mdtraj also requires C++ tools to install and pdbfixer can't be installed via pip (conda-forge instead). Would be great to find a better way to distribute this.

[WillHua127]

i am working on the open-source thing, make everything better!

[pgmikhael]

Hi,

Super exciting work! Will just re-iterate the above. Would also be helpful in general to not import all functions from a script (from file import *) since it makes it harder to follow where things break when/if they do.

[WillHua127]

Thanks Peter! I am cleaning everything. Plan to release everything in Nov.

[WillHua127]

things should be fixed, let me know if you can run enzymeflow_demo.ipynb

[carbondrop-nick]

Thanks! I think we're pretty close, but I am running into a few key naming issues in the pretrain. It looks like ProteinLigandNetwork is expecting some slightly different keys than the ones present:
Missing key(s) in state_dict:
"guide_ligand_mpnn.mpnn.atom_convs.0.lin.weight", "guide_ligand_mpnn.mpnn.atom_convs.1.lin.weight", "guide_ligand_mpnn.mpnn.mol_conv.lin.weight"
Unexpected key(s) in state_dict:
"guide_ligand_mpnn.mpnn.atom_convs.0.lin_src.weight", "guide_ligand_mpnn.mpnn.atom_convs.0.lin_dst.weight", "guide_ligand_mpnn.mpnn.atom_convs.1.lin_src.weight

[pgmikhael]

This may be a matter of the right pyg version that leads to different model implementations. installing torch_geometric with pip and the specified version seems to work.

[carbondrop-nick]

Well spotted! Indeed I was working with a later version of torch_geometric.

[carbondrop-nick]

Looks like
meta_eval_csv = pd.read_csv('data/metadata_eval.csv')
is referring to a flat file that includes absolute paths to Will's computer instead of relative paths:
/Users/willhua/Desktop/EnzymeFlow/data/processed_eval/msa/P07964/P07964.pkl
I edited the file to remove all instances of /Users/willhua/Desktop/EnzymeFlow/ and it seemed to work fine

[WillHua127]

I am working on the multi motif scaffolding, i.e., enzymeflow generate enzyme motifs, I am trying to find the seq_idx or seq_position that maps the motifs back to the whole enzyme. Let me know you have suggestions, or would like to collaborate on this.

[carbondrop-nick]

A good answer would be pretty far beyond me. Enzyme "motif" is a messy concept since catalytic machinery has to be able to access multiple states, making hidden dynamic trajectories important. I would stick to the hidden representation and just try to consider the "foldiness" of the protein separately from its "enzyminess" for a given chemical reaction (enzyminess = 0 for most reactions, >0 for known reaction(s)), but that is much easier said than done.