zaixizhang / flag Goto Github PK

Hello author, thank you very much for your great work. When I run the file “train.py” and process data, there are often cases of skipping, skipping all data. Could you please give me a method to solve that？

Hello author, thank you very much for your great work. May I ask if it is convenient for you to provide your pre-trained model? I saw in the sample.yml that it is the file "./pretrained/model.pt". If you could provide this trained model file, I would be very grateful. Thank you.

We are trying to reproduce the results from the original FLAG paper. We have been able to tweak the training code to make it work, but we still bump into some knotty issues during the sampling/generation stage. Following the original instructions from README.md, we start the sampling process by running the following command:

python motif_sample.py

The Python interpreter gives the following error:

Traceback (most recent call last):
  File "motif_sample.py", line 18, in <module>
    from models.maskfill import MaskFillModel
ModuleNotFoundError: No module named 'models.maskfill'

Then I came to realize: there is no such file as ./models/maskfill.py in the Github repo. I googled for the file and found a file with the same name in the 3DSBDD repo (https://github.com/luost26/3D-Generative-SBDD/blob/main/models/maskfill.py). However, the class __init__() function arguments do not match. In motif_sample.py:412:

    model = MaskFillModel(
        ckpt['config'].model,
      	protein_atom_feature_dim=protein_featurizer.feature_dim,
        ligand_atom_feature_dim=ligand_featurizer.feature_dim,
        vocab=vocab,
        weight=weight,
    ).to(args.device)

The vocab and weight arguments are non-existent in the 3DSBDD version of maskfill.py. I assume the FLAG authors have made substantial changes in the maskfill.py file, but happen not to upload it to Github. I cannot proceed with my experiment reproduction beyond this point before the FLAG version of maskfill.py is uploaded.

Hi Zaixi,

After spending a long time (more than 1 month) reproducing the results reported in your paper, I find that some parts of the codes may be unreasonable.
I will not list every unreasonable one here, but show one example:
https://github.com/zaixizhang/FLAG/blob/main/utils/chemutils.py#L40
The normalization is commented out. But it is obviously necessary here because the direction vector of the rotation axis must be a unit vector. This makes the function not actually physical rotation at all.

After all, the reported performance of this paper is still not reproduced.
Could you provide more information?

During sampling, once the exception of UFF is triggered, an error will come from this line (https://github.com/zaixizhang/FLAG/blob/main/motif_sample.py#L295), and the program will exit.

ValueError: Bad Conformer Id

Could you give some explanations or suggestions to address this?

hi, thank you for nice work.
May i know how you build the dataset files like pdbbind_pocket10_xxx
Would you please share your code?

I ran "python train.py" but it seemed that the training job didn't finish normally. How can I resume the training process with a checkpoint file? Many thanks!

Indexing: 1%|█ | 2200/166398 [00:20<25:42, 106.46it/s]
Traceback (most recent call last):
File "train.py", line 63, in
dataset, subsets = get_dataset(config=config.dataset, transform=transform, )
File "/data/zhouzihan/FLAG-main/utils/datasets/init.py", line 11, in get_dataset
dataset = PocketLigandPairDataset(root, *args, **kwargs)
File "/data/zhouzihan/FLAG-main/utils/datasets/pl.py", line 64, in init
self._precompute_name2id()
File "/data/zhouzihan/FLAG-main/utils/datasets/pl.py", line 132, in _precompute_name2id
data = self.getitem(i)
File "/data/zhouzihan/FLAG-main/utils/datasets/pl.py", line 153, in getitem
data = self.transform(data)
File "/home/ahmu/ENTER/envs/flag_env/lib/python3.8/site-packages/torch_geometric/transforms/compose.py", line 24, in call
data = transform(data)
File "/data/zhouzihan/FLAG-main/utils/transforms.py", line 471, in call
bfs_perm, bfs_focal = self.get_bfs_perm_motif(data['moltree'], self.vocab)
File "/data/zhouzihan/FLAG-main/utils/transforms.py", line 449, in get_bfs_perm_motif
node.wid = vocab.get_index(node.smiles)
File "/data/zhouzihan/FLAG-main/utils/mol_tree.py", line 24, in get_index
return self.vmap[smiles]
KeyError: 'C1CC2CCC(O1)O2'

The error "cands = enum_assemble(self, neighbors)" on line 91 of the mol_tree.py file in the utils folder is occurring because the method enum_assemble is not defined. I also did not find any import statement for this method. Could you please let me know where I can find this method? Thank you.