zaixizhang / mgssl Goto Github PK

Official implementation of NeurIPS'21 paper"Motif-based Graph Self-Supervised Learning for Molecular Property Prediction"

License: Apache License 2.0

Python 100.00%

mgssl's Introduction

Motif-based Graph Self-Supervised Learning for Molecular Property Prediction

Official Pytorch implementation of NeurIPS'21 paper "Motif-based Graph Self-Supervised Learning for Molecular Property Prediction" (https://arxiv.org/abs/2110.00987).

Requirements

pytorch                   1.8.1             
torch-geometric           1.7.0
rdkit                     2020.09.1
tqdm                      4.31.1
tensorboardx              1.6

To install RDKit, please follow the instructions here http://www.rdkit.org/docs/Install.html

motif_based_pretrain/ contains codes for motif-based graph self-supervised pretraining.
finetune/ contains codes for finetuning on MoleculeNet benchmarks for evaluation.

Dataset

For the MoleculeNet dataset for finetuning, we have uploaded them to data.

Training

You can pretrain the model by

cd motif_based_pretrain
python pretrain_motif.py

Evaluation

You can evaluate the pretrained model by finetuning on downstream tasks

cd finetune
python finetune.py

Cite

If you find this repo to be useful, please cite our paper. Thank you.

@article{zhang2021motif,
  title={Motif-based Graph Self-Supervised Learning for Molecular Property Prediction},
  author={Zhang, Zaixi and Liu, Qi and Wang, Hao and Lu, Chengqiang and Lee, Chee-Kong},
  journal={arXiv preprint arXiv:2110.00987},
  year={2021}
}

mgssl's People

Contributors

Stargazers

Watchers

Forkers

ndnng liuchuang0059 jiangjing0122 yifeiwang15 aliang-ai camerondiao seonghaeom biocheming zhwl2117 jongkook-heo gaoshan2006 riccardosabatini shenzhiyang2000 thinhump ku-del

mgssl's Issues

Questions about motifs

Hi, I wonder if there are ovelapping atoms between motifs in one molecule?
I run the following script

from moltree import MolTree
smiles = "O=C1[C@@H]2C=C[C@@H](C=CC2)C1(c1ccccc1)c1ccccc1"
mol = MolTree(smiles)
atom_count = 0
for c in mol.nodes:
    atom_count += c.mol.GetNumAtoms()
    print(c.mol.GetNumAtoms())
    print(c.clique)
print(f"atoms in original mol: {mol.mol.GetNumAtoms()}, atoms in motifs: {atom_count}")

And the output is:

2
[0, 1]
2
[9, 10]
2
[9, 16]
9
[1, 2, 3, 4, 5, 6, 7, 8, 9]
6
[11, 12, 13, 14, 15, 10]
6
[17, 18, 19, 20, 21, 16]
1
[9]
atoms in original mol: 22, atoms in motifs: 28

Why could I get node representation only by GIN?

Hello, Zhang:
When I run the ''model.py'' , only "gin_type" = "gin" work successfully, and other models, including gat, gcn, graphsage and attentivefp are fail to work. All of them were base on your code. I am appreciate you will resoponse to me.

The traceback file is attached.
[Uploading traceback.do
traceback.docx
cx…]()

Best regard,
Xiang Zhang

About Motif-Decomposition

Hi zaixizhang, Thanks for sharing your nice works!

I have some questions about generating cliques and bonds in your code.

Is there any reason to set MST_MAX_WEIGHT to 100?(motif_based_pretrain > util > chemutils.py)
1-1. If not, how does the MST_MAX_WEIGHT effect to generated cliques?(i.e. If MST_MAX_WEIGHT raises, it causes almost infinite number of motifs?)
In MolTree, When the number of edges connecting BRICS cliques is less then 1, Tree decomp overwrites cliques and edges. Does it okay to change the threshold 1 to any small integer like 2 or 3?

Sincerely,
JongKook Heo

Kekulization failed for some atoms.

Thanks for your sharing!
When I run the ''model_tree.py'' ,It has this error"Kekulization failed for some atoms.", and the resulting clique.txt that I run if I expect it is very different from the one you provided.
How can I solve this problem?

How can i get motifs for my own datasets?

A motif that does not exist in the vocabulary：KeyError: 'C1=CNC=N1'

运行pretrain_motif.py后似乎出现不存在于词汇表的motif：

Traceback (most recent call last):
”MGSSL\motif_based_pretrain\pretrain_motif.py", line 221, in <module>
    main()
    ^^^^^^
  File "MGSSL\motif_based_pretrain\pretrain_motif.py", line 214, in main
    train(args, model_list, loader, optimizer_list, device)
  File "MGSSL\motif_based_pretrain\pretrain_motif.py", line 109, in train
    loss, wacc, tacc = motif_model(batch, node_rep)  
                       ^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "torch\nn\modules\module.py", line 1511, in _wrapped_call_impl
    return self._call_impl(*args, **kwargs)
           ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "torch\nn\modules\module.py", line 1520, in _call_impl
    return forward_call(*args, **kwargs)
           ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "MGSSL\motif_based_pretrain\util\motif_generation.py", line 43, in forward
    set_batch_nodeID(mol_batch, self.vocab)
  File "MGSSL\motif_based_pretrain\util\motif_generation.py", line 21, in set_batch_nodeID
    node.wid = vocab.get_index(node.smiles)  
               ^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "MGSSL\motif_based_pretrain\util\mol_tree.py", line 25, in get_index
    return self.vmap[smiles]
           ~~~~~~~~~^^^^^^^^
KeyError: 'C1=CNC=N1'

请问可能是什么原因呢？

[error] rdkit.Chem.rdchem.KekulizeException: Can't kekulize mol.

Hello!
I'm trying to run mol_tree.py, but I'm getting KekulizeException: error.
Is it a problem about rdkit version?
Below are my rdkit version and console output.

rdkit=2021.09.2=py37h13c2175_0

runfile('/home/bionuser/bionlab/bionlab/contrib/seyong/tasks/tmp/MGSSL/motif_based_pretrain/util/mol_tree.py', wdir='/home/bionuser/bionlab/bionlab/contrib/seyong/tasks/tmp/MGSSL/motif_based_pretrain/util')
Python 3.7.12 | packaged by conda-forge | (default, Oct 26 2021, 06:08:53)
Type 'copyright', 'credits' or 'license' for more information
IPython 7.30.1 -- An enhanced Interactive Python. Type '?' for help.
PyDev console: using IPython 7.30.1
Python 3.7.12 | packaged by conda-forge | (default, Oct 26 2021, 06:08:53)
[GCC 9.4.0] on linux
start
CC(C)(C)c1ccc2occ(CC(=O)Nc3ccccc3F)c2c1
C[C@@h]1CC(Nc2cncc(-c3nncn3C)c2)CC@@HC1
N#Cc1ccc(-c2ccc(OC@@Hc3ccccc3)cc2)cc1
CCOC(=O)[C@@h]1CCCN(C(=O)c2nc(-c3ccc(C)cc3)n3c2CCCCC3)C1
N#CC1=C(SCC(=O)Nc2cccc(Cl)c2)N=C([O-])C@HC12CCCCC2
CCNH+C@(CC)C@Hc1cscc1Br
COc1ccc(C(=O)N(C)C@@HC/C(N)=N/O)cc1O
O=C(Nc1nc[nH]n1)c1cccnc1Nc1cccc(F)c1
Cc1c(/C=N/c2cc(Br)ccn2)c(O)n2c(nc3ccccc32)c1C#N
Traceback (most recent call last):
File "/home/bionuser/.conda/envs/bionlab/lib/python3.7/site-packages/IPython/core/interactiveshell.py", line 3457, in run_code
exec(code_obj, self.user_global_ns, self.user_ns)
File "", line 1, in
runfile('/home/bionuser/bionlab/bionlab/contrib/seyong/tasks/tmp/MGSSL/motif_based_pretrain/util/mol_tree.py', wdir='/home/bionuser/bionlab/bionlab/contrib/seyong/tasks/tmp/MGSSL/motif_based_pretrain/util')
File "/home/bionuser/.pycharm_helpers/pydev/_pydev_bundle/pydev_umd.py", line 198, in runfile
pydev_imports.execfile(filename, global_vars, local_vars) # execute the script
File "/home/bionuser/.pycharm_helpers/pydev/_pydev_imps/_pydev_execfile.py", line 18, in execfile
exec(compile(contents+"\n", file, 'exec'), glob, loc)
File "/home/bionuser/bionlab/bionlab/contrib/seyong/tasks/tmp/MGSSL/motif_based_pretrain/util/mol_tree.py", line 149, in
mol = MolTree(line)
File "/home/bionuser/bionlab/bionlab/contrib/seyong/tasks/tmp/MGSSL/motif_based_pretrain/util/mol_tree.py", line 105, in init
cmol = get_clique_mol(self.mol, c)
File "/home/bionuser/bionlab/bionlab/contrib/seyong/tasks/tmp/MGSSL/motif_based_pretrain/util/chemutils.py", line 80, in get_clique_mol
smiles = Chem.MolFragmentToSmiles(mol, atoms, kekuleSmiles=True)
rdkit.Chem.rdchem.KekulizeException: Can't kekulize mol. Unkekulized atoms: 1 2 12 15 23

Thank you!

Are there any released pre-trained models like GROVER?

RDKit ERROR: [07:38:26] Explicit valence for atom # 1 N, 6, is greater than permitted

运行代码：

import sys
sys.path.append('./util/') 
from mol_tree import MolTree

data_mol_tree = []
for idx in range(5):
    smiles = data[idx]
    mol_tree = MolTree(smiles)
    mol_tree.recover()
    mol_tree.assemble()
    data_mol_tree.append(mol_tree)
    mol_tree

报错：
RDKit ERROR: [07:38:26] Explicit valence for atom # 1 N, 6, is greater than permitted
[07:38:26] Explicit valence for atom # 1 N, 6, is greater than permitted
RDKit ERROR: [07:38:26] Explicit valence for atom # 1 N, 6, is greater than permitted
[07:38:26] Explicit valence for atom # 1 N, 6, is greater than permitted
报错指出氮原子 #1 的显式价数为6，超过了允许的最大价数。
请问这个错误，影响后续的预训练和模型的精度性能吗？

Where is the atom loss and bond loss?

Dear author, in the paper the final loss is the sum of atom, bond and motif. I cannot find the loss of atom and bond in the source code. So what is the final loss function of MGSSL?

Error in `python': free(): invalid next size (fast)

感谢你分享的代码，但是在跑bbbp数据集，bace等数据集时出现：
*** Error in `python': free(): invalid next size (fast): 0x00007f10f4028fb0 ***
======= Backtrace: =========
/lib64/libc.so.6(+0x81329)[0x7f13514f1329]
/home/zhuyifeng/miniconda3/envs/MGSSl/lib/python3.9/site-packages/torch/lib/libtorch_cpu.so(_ZN5torch8autograd22DifferentiableViewMetaC1EPN3c1010TensorImplENS2_8optionalINS0_8ViewInfoEEES7_NS0_12CreationMetaE+0x1fc)[0x7f12d864dfdc]
/home/zhuyifeng/miniconda3/envs/MGSSl/lib/python3.9/site-packages/torch/lib/libtorch_cpu.so(+0x2c360fb)[0x7f12d80350fb]
/home/zhuyifeng/miniconda3/envs/MGSSl/lib/python3.9/site-packages/torch/lib/libtorch_cpu.so(+0x2c3695e)[0x7f12d803595e]
/home/zhuyifeng/miniconda3/envs/MGSSl/lib/python3.9/site-packages/torch/lib/libtorch_cpu.so(+0x2c75e6d)[0x7f12d8074e6d]
/home/zhuyifeng/miniconda3/envs/MGSSl/lib/python3.9/site-packages/torch/lib/libtorch_cpu.so(+0x2c761ec)[0x7f12d80751ec]
/home/zhuyifeng/miniconda3/envs/MGSSl/lib/python3.9/site-packages/torch/lib/libtorch_cpu.so(+0x189b239)[0x7f12d6c9a239]
/home/zhuyifeng/miniconda3/envs/MGSSl/lib/python3.9/site-packages/torch/lib/libtorch_cpu.so(_ZNK2at6Tensor1tEv+0x4b)[0x7f12d6c9d2ab]
/home/zhuyifeng/miniconda3/envs/MGSSl/lib/python3.9/site-packages/torch/lib/libtorch_cpu.so(+0x377ea32)[0x7f12d8b7da32]
/home/zhuyifeng/miniconda3/envs/MGSSl/lib/python3.9/site-packages/torch/lib/libtorch_cpu.so(_ZN5torch8autograd9generated13AddmmBackward5applyEOSt6vectorIN2at6TensorESaIS5_EE+0x343)[0x7f12d7f52023]
....
已放弃（吐核）
请问该怎么解决

Run failure(KeyError: 'C1=CSC(C2=NNN=N2)=C1')

hi, when i run this pyhon file:!python pretrain_motif.py
i find some SMILES are not in the clique.txt,how can i fix it and make it run successful?

fingerprint

Hi is there an easy way to extract just the fingerprints given the pretrained model?

Question about dataset

Hey friend, I wanna how to get the dataset?

About kekule

def get_smiles(mol): 
    return Chem.MolToSmiles(mol, kekuleSmiles=True)  
smiles = Chem.MolFragmentToSmiles(mol, atoms, kekuleSmiles=True)

Why does reading smiles require kekule format?
Is it possible not to use kekule format?

Are there some mistakes in tree decomposition modules?

Dear author, I run the mol_tree.py to preprocess the dataset and generate clique.txt, but it throws the KekulizeException for molecules such as Cc1c(/C=C/c2cc(Br)ccn2)c(O)n2c(nc3ccccc32)c1C#N. I use the decomposition modules on other datsets, it make the same exceptions, can you update the code or give some advice?

Pretrained Weights

Hi,

Thank you for uploading the code and the pretrained weights! I have one quick question.

Which mode of MGSSL from the paper does motif_based_pretrain/saved_models/pretrained.pth correspond to?