Accurate prediction of molecular targets using a self-supervised image rep...（代码理解）

GitHub - HongxinXiang/ImageMol: ImageMol is a molecular image-based pre-training deep learning framework for computational drug discovery.

---

Install environment

1. GPU environment

CUDA 10.1

2. create a new conda environment

conda create -n imagemol python=3.7.3

conda activate imagemol

3. download some packages

conda install -c rdkit rdkit

windows:

• pip install https://download.pytorch.org/whl/cu101/torch-1.4.0-cp37-cp37m-win_amd64.whl
• pip install https://download.pytorch.org/whl/cu101/torchvision-0.5.0-cp37-cp37m-win_amd64.whl

linux:

• pip install https://download.pytorch.org/whl/cu101/torch-1.4.0-cp37-cp37m-linux_x86_64.whl
• pip install https://download.pytorch.org/whl/cu101/torchvision-0.5.0-cp37-cp37m-linux_x86_64.whl

pip install torch-cluster torch-scatter torch-sparse torch-spline-conv -f https://pytorch-geometric.com/whl/torch-1.4.0%2Bcu101.html

pip install -r requirements.txt

source activate imagemol

Pretraining

1. preparing dataset

Download pretraining data and put it into ./datasets/pretraining/data/

Preprocess dataset:（这个时间比较长）

python ./data_process/smiles2img_pretrain.py --dataroot ./datasets/pretraining/ --dataset data

Note: You can find the toy dataset in ./datasets/toy/pretraining/

2. start to pretrain

Usage:

usage: pretrain.py [-h] [--lr LR] [--wd WD] [--workers WORKERS]
                   [--val_workers VAL_WORKERS] [--epochs EPOCHS]
                   [--start_epoch START_EPOCH] [--batch BATCH]
                   [--momentum MOMENTUM] [--checkpoints CHECKPOINTS]
                   [--seed SEED] [--dataroot DATAROOT] [--dataset DATASET]
                   [--ckpt_dir CKPT_DIR] [--modelname {ResNet18}]
                   [--verbose] [--ngpu NGPU] [--gpu GPU] [--nc NC] [--ndf NDF]
                   [--imageSize IMAGESIZE] [--Jigsaw_lambda JIGSAW_LAMBDA]
                   [--cluster_lambda CLUSTER_LAMBDA]
                   [--constractive_lambda CONSTRACTIVE_LAMBDA]
                   [--matcher_lambda MATCHER_LAMBDA]
                   [--is_recover_training IS_RECOVER_TRAINING]
                   [--cl_mask_type {random_mask,rectangle_mask,mix_mask}]
                   [--cl_mask_shape_h CL_MASK_SHAPE_H]
                   [--cl_mask_shape_w CL_MASK_SHAPE_W]
                   [--cl_mask_ratio CL_MASK_RATIO]

Code to pretrain:

python pretrain.py --ckpt_dir ./ckpts/pretraining/ \
                   --checkpoints 1 \
                   --Jigsaw_lambda 1 \
                   --cluster_lambda 1 \
                   --constractive_lambda 1 \
                   --matcher_lambda 1 \
                   --is_recover_training 1 \
                   --batch 256 \
                   --dataroot ./datasets/pretraining/ \
                   --dataset data \
                   --gpu 0,1,2,3 \
                   --ngpu 4

For testing, you can simply pre-train ImageMol using single GPU on toy dataset:

python pretrain.py --ckpt_dir ./ckpts/pretraining-toy/ \
                   --checkpoints 1 \
                   --Jigsaw_lambda 1 \
                   --cluster_lambda 1 \
                   --constractive_lambda 1 \
                   --matcher_lambda 1 \
                   --is_recover_training 1 \
                   --batch 16 \
                   --dataroot ./datasets/toy/pretraining/ \
                   --dataset data \
                   --gpu 0 \
                   --ngpu 1