hicai-zju / imold

Official implementation for Learning Invariant Molecular Representation in Latent Discrete Space (NeurIPS 2023)

Learning Invariant Molecular Representation in Latent Discrete Space

This repository is the official implementation of our paper:

Learning Invariant Molecular Representation in Latent Discrete Space

Xiang Zhuang, Qiang Zhang*, Keyan Ding, Yatao Bian, Xiao Wang, Jingsong Lv, Hongyang Chen, Huajun Chen* (* denotes correspondence)

Advances in Neural Information Processing Systems (NeurIPS) 2023

Environment

To run the code successfully, the following dependencies need to be installed:

Python                     3.8      
torch                      1.10.1
torch_geometric            2.0.4
torch_scatter              2.0.9
torch_cluster              1.6.0
torch_sparse               0.6.13
torch_spline_conv          1.2.1
rdkit_pypi                 2022.9.5
vector_quantize_pytorch    1.0.7
ogb                        1.3.6

This repo is also depended on GOOD and DrugOOD, please follow the installation methods provided for each package:

• GOOD (Version 1.1.1)
  • Repository: https://github.com/divelab/GOOD/
  • Installation: Please follow the instructions provided in the repository to install.
• DrugOOD (Version 0.0.1)
  • Repository: https://github.com/tencent-ailab/DrugOOD
  • Installation: Please follow the instructions provided in the repository to install.

Data

The data used in the experiments can be downloaded from the following sources:

1. GOOD
   • GOODPCBA
   • GOODHIV
   • GOODZINC
   • Extract the downloaded files and save the contents in the data directory.
2. DrugOOD
   • download from link.
   • Extract the downloaded file and save the contents in the drugood-data-chembl30 directory.

An example of the folder hierarchy after adding the data files:

├── data
│   ├── GOODHIV
│   ├── GOODPCBA
│   ├── GOODZINC
├── drugood-data-chembl30
│   ├── lbap_core_ec50_assay.json
│   └── ...
├── models
│   ├── model.py
│   └── ...
├── run.py
└── README.md

Running Script

Training

python run.py --dataset GOODZINC --domain scaffold --shift concept --num_e 4000 --bs 256 --gamma 0.5 --inv_w 0.01 --reg_w 0.5 --gpu 0 --exp_name ZINC --exp_id scaffold-concept

Running parameters and descriptions are as follows:

Parameter	Description	Choices
dataset	name of dataset	GOODHIV, GOODZINC, GOODPCBA, ic50_assay, ic50_scaffold, ic50_size, ec50_assay, ec50_scaffold, ec50_size.
domain	environment-splitting strategy	scaffold, size. Only need to be specified for datasets in GOOD.
shift	type of distribution shift	covariate, concept. Only need to be specified for datasets in GOOD.
num_e	code book size	-
bs	batch size	-
gamma	threshold $\gamma$	-
inv_w	$\lambda_1$	-
reg_w	$\lambda_2$	-
gpu	which GPU to use	-
exp_name	experiment name	-
exp_id	experiment ID	-

Evaluation

We provide the hyperparameters for the training of each dataset in the Appendix, and provide the corresponding checkpoints in the release page.

python eval.py --dataset GOODZINC --domain scaffold --shift concept --load_path checkpoint/GOODZINC-scaffold-concept.pkl

The load_path parameter specifies the path to load the checkpoint.

Citation

If you use or extend our work, please cite the paper as follows:

@InProceedings{zhuang2023learning,
  title={Learning Invariant Molecular Representation in Latent Discrete Space},
  author={Xiang Zhuang and Qiang Zhang and Keyan Ding and Yatao Bian and Xiao Wang and Jingsong Lv and Hongyang Chen and Huajun Chen},
  booktile={Advances in Neural Information Processing Systems},
  year={2023}
}

imold's Issues

Can not reproduce the result on DrugOOD.

I ran your code on DrugOOD with 10 different seeds (0 ~ 9), but the results are significantly lower than those reported in your paper. For example, the ROC-AUC on ec50_assay and ec50_size are:

• ec50_assay: 76.997 73.57 71.06 72.65 75.16 73.07 71.84 71.92 73.67 76.86
• ec50_size : 61.71 61.93 60.12 61.95 62.50 61.92 57.81 62.97 61.85 64.20

these results are all below those reported in your paper.
I have used the same hyper-parameters provided in the Appendix. Does that mean your proposed method is sensitive to the random seed?
Thanks!

requests.exceptions.ConnectionError:

requests.exceptions.ConnectionError: HTTPSConnectionPool(host='drive.google.com', port=443): Max retries exceeded with url: /uc?id=1GNc0HUee5YQH4Vtlk8ZbDjyJBYTEyabo (Caused by NewConnectionError('<urllib3.connection.HTTPSConnection object at 0x7fc3f5ebb0a0>: Failed to establish a new connection: [Errno 110] Connection timed out'))

Can you tell me how to solve it?

Running Error

Dear Authors,

When I run this code, there exists an error. Could you give me some advice?

About Siam Projection

iMoLD/models/model.py

Line 93 in 6d24452

self.simsiam_proj = nn.Sequential(nn.Linear(emb_dim, emb_dim * 2),

Thanks for sharing the codes!

But I am confused about one part: It seems that the projection layer in the Encoder isn't used anywhere in the code. Did I miss something?

About Evaluation

Dear Authors,

I have finished the training process. However, I cannot find the saved model.

Is the model saved in dump/0813-ZINC/scaffold-concept/params.pkl?

But when I run the command, these exists an error, i.e., python eval.py --dataset GOODZINC --domain scaffold --shift concept --load_path dump/0813-ZINC/scaffold-concept/params.pkl?