josejimenezluna / delfta Goto Github PK

For completeness, as we say we provide those in the manuscript.

Hi! Thanks for a great repository.
I am trying to download the files as suggested in the README, but unfortunately I get this error:

(delfta) mduranfrigola@raluy:~/Desktop$ python -c "import runpy; _ = runpy.run_module('delfta.download', run_name='__main__')"
2024/04/24 05:17:59 PM | DelFTa | INFO: Now downloading trained models and utils...
Traceback (most recent call last):                                                                                                  
  File "<string>", line 1, in <module>
  File "/home/mduranfrigola/miniconda3/envs/delfta/lib/python3.9/runpy.py", line 228, in run_module
    return _run_code(code, {}, init_globals, run_name, mod_spec)
  File "/home/mduranfrigola/miniconda3/envs/delfta/lib/python3.9/runpy.py", line 87, in _run_code
    exec(code, run_globals)
  File "/home/mduranfrigola/miniconda3/envs/delfta/lib/python3.9/site-packages/delfta/download.py", line 134, in <module>
    _download_required()
  File "/home/mduranfrigola/miniconda3/envs/delfta/lib/python3.9/site-packages/delfta/download.py", line 93, in _download_required
    with tarfile.open(models_tar) as handle:
  File "/home/mduranfrigola/miniconda3/envs/delfta/lib/python3.9/tarfile.py", line 1797, in open
    raise ReadError("file could not be opened successfully")
tarfile.ReadError: file could not be opened successfully

Do you know what might be going on?
Thanks a lot in advance.

When using the docker image from DockerHub, the import of openbabel fails:

>>> from openbabel.pybel import readstring
==============================
*** Open Babel Error  in openLib
  /opt/conda/envs/delfta/lib/openbabel/3.1.0/png2format.so did not load properly.
 Error: libXau.so.6: cannot open shared object file: No such file or directory

This can be solved by running (inside the container):

apt-get update
apt-get install libxtst6

(see also https://stackoverflow.com/questions/17355863/cant-find-install-libxtst-so-6)

Hi, I would like to use the QMugs dataset used in this work with PyTorch Geometric. I am trying to set it up as an InMemoryDataset similar to how QM9 has been setup in PyG: https://pytorch-geometric.readthedocs.io/en/latest/_modules/torch_geometric/datasets/qm9.html#QM9

Could you please guide on converting QMugs to a PyG dataset? for eg., which urls should I use for raw_url and raw_dir ?

• make3D automatically assigns hydrogens, so when passing a molecule without hydrogens but using force3D=True, there's no feedback in the logger that hydrogens have been added. Additionally, this means that force3d=True overrides addh=False. I suggest first checking if hydrogens are present, and, if that is not the case but force3d=True, to return an error.
• Checking for hydrogens with 1 in set([atom.atomicnum for atom in mol.atoms]) runs the risk of ignoring molecules where only certain hydrogens are present. It's a strange edge case, but I suggest checking if e.g. [atom.atomicnum for atom in mol.atoms] changes after adding hydrogens to a copy of the molecule. Not elegant, but didn't find an alternative yet.

to provide meV errors in the manuscript

• need to download to and run from the same folder
• assert len(mols) == 100 in test_xtb.py and test_calculator.py to make sure we actually have something to run the tests on

I have installed delfta with conda install delfta -c delfta -c pytorch -c rusty1s -c conda-forge on a WSL Ubuntu 20.04
Now I'm trying to execute the "First Run" commands, but both
python -c "import runpy; _ = runpy.run_module('delfta.download', run_name='__main__')" and

from delfta.download import _download_required
_download_required()

throw tarfile.ReadError: file could not be opened successfully
The file in question is under home/user/miniconda3/envs/delfta/lib/python3.8/tarfile.py

Having an option to set the logger level (e.g., suppress all INFO-level messages) would be nice.

When an invalid input file is provided, OpenBabel does print a warning, but we try to keep going and return somewhat confusing errors.

• example_file_1.sdf returns Molecules at position [0] have no 3D conformations available. Either provide a mol with one or re-run calculator with force3D=True.
• example_file_2.sdf returns need at least one array to concatenate

example_files.zip

Checking if molecules are valid and throwing an error if not might be clearer for the user. In the two cases above, input_ is either a list of empty (no atoms) pybel molecules (for example_file_1.sdf) or just an empty list (for example_file_2.sdf), so we could check for that.

Instead of building the container each time with the Dockerfile. TBD once repo goes public

When running DelftaCalculator for more than batch_size molecules, new progress bars are created for each batch. This somewhat defeats the purpose of a progress bar, as it always only goes from 0/1 to 1/1, and then a new bar is created, so there's no estimate on how long the overall process will take, or which percentage of the overall task has been completed. As a minor point, we can probably skip the progress if the input is < batch_size, although the user could also turn that off themselves.

Python 3.7 is approaching EOL soon.

e.g. smiles, inchi

This error (

To see if error propagation still works correctly.

• mol.addhs()
• check for atom types that are not included in QMugs

mol.atoms can only be accessed if the previous check passed for all items in the list.

Useful when computing a large number of molecules.

It is quite ugly to expose this parameter to the user.
Alternatively, consider making this a class variable as in self.offset_idx.

For large numbers of molecules (i.e. an sdf file) it may make more sense to use joblib with large batch sizes.

This is more of a thought than a bug issue, but couldn't it all go to the calculator.py file?

Using a modern GPU

Combinations of force3d with addh can lead to unexpected results.

While I realize E_gap is a separate model itself, the value it returns is sometimes way different from calculating E_lumo - E_homo. For example, take the following molecule with the following settings:

calc = DelftaCalculator(delta=True, xtbopt=True, return_optmols=True, addh=True)
mol = readstring("smi", "C[C@@H]1CC[C@@H]2C(C)(C)[C@H](O)[C@]3(C)CC[C@@]21C3")

Results are:

'E_homo': -0.23676376
'E_lumo': 0.11727003
'E_gap': 0.12012529

E_gap should be close to 0.354

via conda

xtb