SCAGE

What is SCAGE?

SCAGE is a self-conformation-aware pre-training framework for molecular property prediction reveals the quantitate structure-activity relationship like human experts

Reference paper

Geometry-enhanced molecular representation learning for property prediction
Accurate prediction of molecular properties and drug targets using a self-supervised image representation learning framework

Some command lines

You can directly use the following command lines to run the code in bash script.

CUDA_VISIBLE_DEVICES=1 nohup python pretrain.py > ./pretrain_log/pretrain1.log 2>&1 &

0. Fix some bugs

export LD_LIBRARY_PATH="/home2/s439850/anaconda3/envs/SAGE/lib:$LD_LIBRARY_PATH"

export LD_LIBRARY_PATH="/home2/s439850/anaconda3/envs/SAGEHH/lib:$LD_LIBRARY_PATH"

1. Pre-training

CUDA_VISIBLE_DEVICES=3 python pretrain.py

python -m torch.distributed.launch --nproc_per_node=4 pretrain_dis.py

2. Fine-tuning

unset http_proxy
unset https_proxy

nnictl create --config ./config.yaml -p 3325

experimental results

tencent document link

close nni and then view it again

nnictl stop --port 3329
nnictl stop -a

nnictl view omFiEsra --port 3328 -e /archive/bioinformatics/Zhou_lab/shared/jjin/SCAGE/log

1. fintune on no pretrain

some other results

2. fintune on no pretrain with pretrain parameters

Some conclusions:

Choose dist bar as [1,2,3]
Choose dropout to 0.3/0.4
Choose number encoder 4
Choose number layer 4

FakeEnd / SCAGE