Skip to content

Latest commit

 

History

History
90 lines (71 loc) · 4.25 KB

dp.md

File metadata and controls

90 lines (71 loc) · 4.25 KB
Raw

Drug Property Prediction

Drug property prediction (DP) aims to predict molecule properties such as toxicity and side effects, which is a significant task in drug discovery.

Features

  • Supported models: MolCLR, GraphMVP, MoMu, MolFM and DeepEIK. This is a continuing effort and we are working on further growing the list.
  • Supported datasets: 8 classification datasets i.e. BBBP, Tox21, ToxCast, SIDER, ClinTox, MUV, HIV and BACE of MoleculeNet.
  • Supported split: random split, scaffold split and random-scaffold split;
  • Supproted evaluation: ROC-AUC.

Data Preparation

Download MoleculeNet datasets here, unzip the file, and put the dataset fold under datasets/dp/. You can use the following commands from within OpenBioMed/:

wget http://snap.stanford.edu/gnn-pretrain/data/chem_dataset.zip
unzip chem_dataset.zip
mkdir -p datasets/dp
mv dataset datasets/dp/moleculenet
rm chem_dataset.zip

After downloading and unzipping, you should remove all the processed/ directories of 8 datasets in the dataset/ folder. Otherwise you will get the following error:

RuntimeError: The 'data' object was created by an older version of PyG. If this error occurred while loading an already existing dataset, remove the 'processed/' directory in the dataset's root folder and try again.

Model Preparation

To reproduce DeepEIK, you should install PubMedBERT (uncased) from huggingface and put the checkpoint under ckpts/text_ckpts/.

To reproduce or finetune MolCLR, MoMu and GraphMVP used their trained checkpoints, you can download the checkpoints from:
MolCLR: https://github.com/yuyangw/MolCLR
MoMu: https://github.com/ddz16/MoleculePrediction
GraphMVP: https://github.com/chao1224/GraphMVP

You will need to rename and put the checkpoints at the following paths.

# MolCLR
ckpts/gnn_ckpts/molclr/model.pth
# Momu
ckpts/fusion_ckpts/momu/MoMu-K.ckpt
# GraphMVP
ckpts/gnn_ckpts/graphmvp/pretraining_model.pth

Training and Evaluation

You can run scripts using bash under scripts/aidd/dp/:

scripts/aidd/dp
├── train_molclr.sh		 # running MolCLR on 8 datasets of moleculenet
├── train_graphmvp.sh  # running GraphMVP on 8 datasets of moleculenet
├── train_momu.sh      # running MoMu on 8 datasets of moleculenet
├── train_molfm.sh     # running MolFM on 8 datasets of moleculenet
└── train_deepeik.sh   # running DeepEIK on 8 datasets of moleculenet

Example:

bash scripts/aidd/dp/train_molfm.sh cuda:0   # switch to your on cuda device or cpu

You can also modify the scripts or directly use the following command:

python open_biomed/tasks/mol_task/dp.py \
[--device DEVICE] \									  # gpu device id
[--mode MODE] \											  # training mode, train: train-test
[--config_path CONFIG_PATH] \				  # configuration file, see configs/dp/ for more details
[--dataset DATASET] \								  # datasets name, support MoleculeNet now
[--dataset_path DATASET_PATH] \       # path to the datasets
[--dataset_name DATASET_NAME] \       # name of the dataset
[--init_checkpoint INIT_CHECKPOINT] \ # checkpoint path used for efficient validation
[--param_key PARAM_KEY] \							# key of the checkpoint dict that contains model parameters
[--output_path OUTPUT_PATH] \         # save checkpoint path used for training
[--num_workers NUM_WORKERS] \         # number of workers when loading data
[--patience PATIENCE] \               # number of tolerant epochs for early-stopping
[--weight_decay WEIGHT_DECAY] \       # weight decay, default is 1e-5
[--lr LR] \                           # learning rate, default is 1e-3
[--batch_size BATCH_SIZE] \           # batch size, default is 128
[--epochs EPOCHS] \                   # number of training epochs
[--logging_steps LOGGING_STEPS] \     # steps for printing training information
[--seed SEED]                         # random seed
[--dropout DROPOUT]                   # The dropout ratio of dp model