openbabel processing error with crossdocked data

[SanazKaz]

Hi all,

Currently trying to pre-process the data using process_crossdocked.py.

i receive this #failed: 89: 100%|██████████| 100000/100000 [29:33<00:00, 56.39it/s]

i have 2 questions -

1. is it expected to have 89 failed files? I thought all 100,000 should pass fine.

2. While it is processing files for training i received this error below

• There are actually quite a few openbabel warnings as the files are being processed.

---

*** Open Babel Warning in PerceiveBondOrders
Failed to kekulize aromatic bonds in OBMol::PerceiveBondOrders (title is /tmp/tmpn058zqxm)

[16:10:15] Unexpected error hit on line 60
[16:10:15] ERROR: moving to the beginning of the next molecule

7352/7365 successful: 7%|▋ | 7365/99911 [00:20<04:11, 368.12it/s]
Traceback (most recent call last):
File "/data/stat-cadd/wolf7055/diffsbdd-ppo/process_crossdock.py", line 422, in
train_smiles = compute_smiles(lig_coords, lig_one_hot, lig_mask)
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "/data/stat-cadd/wolf7055/diffsbdd-ppo/process_crossdock.py", line 138, in compute_smiles
mol = build_molecule(pos, atom_type, dataset_info)
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "/data/stat-cadd/wolf7055/diffsbdd-ppo/analysis/molecule_builder.py", line 154, in build_molecule
mol = make_mol_openbabel(positions, atom_types,
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "/data/stat-cadd/wolf7055/diffsbdd-ppo/analysis/molecule_builder.py", line 90, in make_mol_openbabel
for atom in tmp_mol.GetAtoms():
^^^^^^^^^^^^^^^^
AttributeError: 'NoneType' object has no attribute 'GetAtoms'

---

Thanks in advance!

[SanazKaz]

Error in file process_crossdocked.py with incorrect dict - see below for fix.

• although this still results in 89 failed mols due to sanitisation, at least the script runs.

def process_ligand_and_pocket(...):
...
try:
            pocket_one_hot = []
            for a in full_atoms:
                if a in atom_dict: ################################### previously amino_acid_dict 
                    atom = np.eye(1, len(atom_dict),
                                  atom_dict[a.capitalize()]).squeeze()
                elif a != 'H':
                    atom = np.eye(1, len(atom_dict),
                                  len(atom_dict)).squeeze()
                pocket_one_hot.append(atom)
            pocket_one_hot = np.stack(pocket_one_hot)
        except KeyError as e:
            raise KeyError(
                f'{e} not in atom dict ({pdbfile})')
        pocket_data = {
            'pocket_coords': full_coords,
            'pocket_one_hot': pocket_one_hot,
            'pocket_ids': pocket_ids
        }
    return ligand_data, pocket_data

[charlotte0104]

hi, I encountered the same issue. Could you replicate the results of the paper using your own processed dataset?

[SanazKaz]

Hi Charlotte, i did not manage to get the mols to 100k, although i suspect this is due to some hydrogen issue which can be fixed with some messing around. (i may be wrong here, but this is what seems to cause the failure)

I have carried on with the rest of my code with the 89 failed mols. For my purposes this is fine.

[charlotte0104]

Thank you for your reply. I found that during data processing, there were cases where the number of ligand atoms and pocket atoms did not match in the data being stored in the dataset, which caused errors in the subsequent data. I avoided this issue by ignoring the erroneous data and then started the training. However, when training DiffSBDD, the training pauses for about five seconds after every fifty batches and then resumes. And I have not been able to identify the cause of the issue. Have you encountered this issue? Do you have any good solutions or suggestions?
I would be very grateful to receive your next reply

[charlotte0104]

thanks for all your help and i have found why the training pauses for about five seconds after every fifty batches. the ligntning model writes loss and other log files to the disk every 50 batches, which significantly impacts the training speed.
Commenting out #self.log_metrics(info, 'train', batch_size=len(data['num_lig_atoms'])) in the training_step resolves the issue.

[SanazKaz]

hi Charlotte! Great, i'm glad you solved it. I am actually playing around with the checkpoint models.

Checkout my other issue i opened for the mismatch between the data processing (namely, the "other" category when encoding atoms) for the issue with the data encoding mismatch (temporary) fix.

[mljchen102]

I'm getting a similar error but when I tried your fix, it didn't work. What else can cause this issue?

Computing SMILES ...
26/26 successful: 0%| | 0/764 [00:00<?, ?it/s]==============================
*** Open Babel Warning in PerceiveBondOrders
Failed to kekulize aromatic bonds in OBMol::PerceiveBondOrders (title is /tmp/tmp3d0s4ad0)

65/65 successful: 0%| | 0/764 [00:00<?, ?it/s][10:16:17] Unexpected error hit on line 73
[10:16:17] ERROR: moving to the beginning of the next molecule
65/65 successful: 9%|██████████████████▌ | 65/764 [00:00<00:00, 893.92it/s]
Traceback (most recent call last):
File "/home/michael/DiffSBDD/process_cb1_antag_data.py", line 401, in
train_smiles = compute_smiles(lig_coords, lig_one_hot, lig_mask)
File "/home/michael/DiffSBDD/process_cb1_antag_data.py", line 141, in compute_smiles
mol = build_molecule(pos, atom_type, dataset_info)
File "/home/michael/DiffSBDD/analysis/molecule_builder.py", line 154, in build_molecule
mol = make_mol_openbabel(positions, atom_types,
File "/home/michael/DiffSBDD/analysis/molecule_builder.py", line 90, in make_mol_openbabel
for atom in tmp_mol.GetAtoms():
AttributeError: 'NoneType' object has no attribute 'GetAtoms'