This repository provides source code, and data associated with our Journal of Chemical Information and Modeling publication, "Large-Scale Pretraining Improves Sample Efficiency of Active Learning-Based Virtual Screening".
Paper: JCIM Link / Arxiv Link
PretrainedAL-VS is an Active learning framework with the pretained large Language model added as surrogate model.
Our code is based on MolPal and MolFormer, please check the installation of MolPal and MolFormer.
- MolPal: https://github.com/coleygroup/molpal
- MolFormer: https://github.com/IBM/molformer
The model training, inference, and molecular docking parts are customized to run on HPC via SLURM RESTAPI. Please check the slide for the detail (link)
Download pretrained MolFormer model from the link, and put it in molpal/models/transformer/pretrained_ckpt/
Config slurm API server for model training and inference in molpal/models/transformermodels.py
def __init__(
self,
n_iter: int = 0,
uncertainty: Optional[str] = 'none',
ngpu: int = 1,
ddp: bool = False,
slurm_token: Optional[str] = None,
log_dir: Optional[Union[str, Path]] = None,
work_dir: Optional[Union[str, Path]] = None,
weight_path: str = 'molpal/models/transformer/pretrained_ckpt/pretrained_weights.ckpt',
seed: Optional[int] = None
):
self.n_iter = n_iter
self.CPU_PER_GPU = 8
self.seed_path = os.path.join(work_dir, weight_path)
self.ngpu = ngpu
self.ncpu = ngpu*self.CPU_PER_GPU
self.n_node, self.tasks_per_node = self.get_n_node(ngpu)
self.ddp = ddp
self.log_dir = log_dir
self.seed = seed
self.work_dir = work_dir
self.slurm_url = 'Your Slurm Management Node ULR and port'
self.user_name = os.environ.get("SLURM_USER", getpass.getuser())
self.version = 'v0.0.38'
self.errorTolerance = 20
if slurm_token == None:
self.slurm_token = self.get_slurm_token(days=1)
self.uncertainty = uncertaintyConfig slurm API server for SCHRODINGER/glide docking in molpal/objectives/glide.py if you want to use SCHRODINGER glide
def __init__(
self,
objective_config: str,
minimize: bool = True,
**kwargs,
):
self.slurm_url = 'http://edge-hpc-mgt-101.hpc.biogen.com:6820'
self.user_name = os.environ.get("SLURM_USER", getpass.getuser())
self.slurm_token = self.get_slurm_token(days=1)
self.version = 'v0.0.38'
ncpu, utils_dir, targetfname = parse_config(objective_config)
self.utils_dir = utils_dir
# self.ncpu = 20
self.ncpu = ncpu
self.targetfname = targetfname
self.errorTolerance = 20
super().__init__(minimize=minimize)
self.word_pair = {
'UTILSPATH': self.utils_dir,
'TARGETFNAME': self.targetfname
} sbatch run_molpal.sh
Ye Wang ([email protected])