Protseed-Cath

Environment

We provide a conda environment file scripts/environment.yml for Linux machine with CUDA 11.3 / 11.4 installed. Pleaes modify the environment.yml if you want to use other CUDA version. To install the environment, run

bash scripts/install_third_party_dependencies.sh

Replace the mamba with conda in the script if you do not use mambaforge.

Checkpoints and Data

We provide the checkpoints and data used in the paper in Google Drive. Please download the checkpoints and put them in the checkpoints folder. The md5sum of the checkpoints and data are listed in scripts/md5sum.txt.

If you want to train your own model, please modify the data path in scripts/train.sh.

Evaluation

Please see scripts/eval.sh for the evaluation script.

python run_cath.py \
    $TEST_DIR/3tm4A01.pdb \
    $SCRATCH/datasets/structure_datasets/cath/raw/ss_annotation_31885.pkl \
    $ckpt_path \
    --yaml_config_preset yaml_config/deterministic.yml \
    --output_dir $SCRATCH/projects_output/cath_gen/inference/eval \
    --model_device cuda:0 \
    --no_recycling_iters 3 \
    --deepspeed false \
    --relax false \
    --seed 42

python batch_run_cath.py \
    $TEST_DIR \
    $SCRATCH/datasets/structure_datasets/cath/raw/ss_annotation_31885.pkl \
    $ckpt_path \
    --yaml_config_preset yaml_config/deterministic.yml \
    --output_dir $SCRATCH/projects_output/cath_gen/inference/eval/deterministic \
    --model_device cuda:0 \
    --no_recycling_iters 3 \
    --relax false \
    --deepspeed false \
    --seed 9141423

Training

Please see scripts/train.sh for the training script.

python train_cath.py $TRAIN_DIR $OUTPUT_DIR \
    --ss_file $SCRATCH/datasets/structure_datasets/cath/raw/ss_annotation_31885.pkl \
    --val_data_dir $VALID_DIR \
    --seed 2022 \
    --yaml_config_preset yaml_config/deterministic.yml \
    --precision 16 --gpus 4 --log_every_n_steps 50 \
    --wandb true \
    --wandb_entity chenceshi \
    --wandb_version release \
    --wandb_project cath_gen \
    --train_epoch_len 2000 \
    --gradient_clip_val 1.0

Note that we use wandb to record the training process. Please modify the wandb_entity and wandb_project if you want to use your own wandb account.

Protseed-Antibody Generation

Coming very soon!