The code is currently organized to be run in 3 separate chunks. Tutorial below

is for using the command line interface. For more custom solutions, interface

with the code directly (examples are in the 'example_api_scripts' dir.

(i.e. data_processing_submit.py, train_submit.py, and analysis_submit.py)

The first step is to convert raw trajectories into diffnet input, which

should be performed on a CPU node.

python /path/to/diffnets/diffnets/cli/main.py process {sim_dirs} {pdb_fns} {outdir}

where {sim_dirs} is a path to an np.array containing directory names. The array

needs one directory name (string) for each variant where each directory contains

all trajectories for that variant. {pdb_fns} is a path to an np.array containing

pdb filenames. The array needs one pdb filename for each variant. The order of

variants should match the order of {sim_dirs}. {outdir} is the path you would

like processed data to live. Examples of these files can be found in example_cli_files.

You can optionally include -a {atom-sel} where {atom-sel} is a path to an

np.array containing a list of indices for each variant, which operates on the

pdbs supplied. The indices need to select equivalent atoms across variants.

Without specifying tatom_sel, the default will grab C, CA, CB, and N atoms

from each variant. If this doesn't in an equivalent number of atoms across

variants, the code will fail and you will need to specify atom_sel. Additionally,

you can optionally include -s {stride} where stride is a path to a np.array

containing a list of integers that correspond to how much trajectory subsampling

you want to perform for each variant. Examples of these files can be found in

example_cli_files.

Next, train a DiffNet with the folowing command:

python /path/to/diffnets/diffnets/cli/main.py train config.yml

where config.yml contains all the training parameters. Look at train_sample.yml

as an example and train_sample.txt for descriptions of each parameter. Training

on a GPU gives better performance than on a CPU.

Finally, run some automated analyses with...

python /path/to/diffnets/diffnets/cli/main.py analyze {data_dir} {net_dir}

where {data_dir} is the path to the directory output by the earlier 'process'

command, and {net_dir} is the path to the directory output by the earlier 'train' command.

This analysis includes reconstructing all trajectories using the DiffNet,

calculating an RMSD between DiffNet reconstructed structures and their respective

simulation frame, calculating classification labels for all frames, and

calculating the latent vector for all frames. Additionally, this script is setup

to generate a .pml file in the {net_dir}. Loading {data_dir}/master.pdb into

pymol followed by loading this .pml file will generate a figure like Figure 7 in

the paper. Finally, this will provide a `morph` directory that contains a PDB

containing 10 structures the represent DiffNet labels changing from 0 to 1.