ProtcdNl

Datasets and Checkpoint

The dataset and checkpoint is available on Google Drive:

Download

Environment

We provide a conda environment file scripts/environment.yml for Linux machine with CUDA 11.3 installed. Please modify the environment.yml if you want to use other CUDA version.

Evaluation

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

python run_cath.py \
    $TEST_DIR/3tm4A01.pdb \
    $SCRATCH/datasets/structure_datasets/cath/raw/all_combined_info.json \
    $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/all_combined_info.json \
    $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/all_combined_info.json \
    --val_data_dir $VALID_DIR \
    --seed 2025 \
    --yaml_config_preset yaml_config/deterministic.yml \
    --precision 16 --gpus 4 --log_every_n_steps 50 \
    --wandb true \
    --wandb_entity xxx \
    --wandb_version protcdnl \
    --wandb_project protcdnl \
    --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.