Our code works on Linux well with CUDA==11.4. It will be required to install pytorch and other dependencies listed below with the corresponding CUDA version (if necessary) on your server.
biopython
e3nn
pandas
torch_scatter
tqdm
requests
The raw DB5.5 dataset was already placed in the data directory from the original source:
https://zlab.umassmed.edu/benchmark/ or https://github.com/drorlab/DIPS
The raw PDB files of DB5.5 dataset are in the directory ./data/benchmark5.5/structures, with train/val/test.txt in the directory ./data/benchmark5.5 as the exact dataset split for use.
We use SAbDab as the training set and RAbD as the test set. SAbDab is a database of antibody structures that updates on a weekly basis, and RAbD is a curated benchmark on antibody design. We have provided the summary data for SAbDab and RAbD in the summaries folder, please further download all structure data (renumbered with IMGT in PDB format) from the downloads entry of the offical website of SAbDab. Ensure the structure data renumbered by imgt is located at {PDB_DIR}/all_structures/imgt, while PDB_DIR is up to you to choose.
We have provided the script in data/download_and_split.sh to preprocess and split the data of SAbDab. Please run the following command to get your data prepared:
cd data && bash download_and_split.sh -d {PDB_DIR}If all goes well, you will see train/valid/test.json in the PDB_DIR directory.
The training bash file train_db5.sh and train_sabdab.sh are given. You can begin your training process as below:
# for DB5.5
GPU=0 bash train_db5.sh
# for SAbDab
GPU=0 bash train_sabdab.shyou must change DATA_DIR in train_sabdab.sh, which should be your own PDB_DIR. You can change other hyperparameters for training in *.sh files.
We have provided the source cpp code TMscore.cpp in evaluate, please run:
cd evaluate && g++ -O3 -o TMscore TMscore.cppWe use the public tool from Github (https://github.com/bjornwallner/DockQ) to evaluate DockQ, please run:
cd evaluate && git clone https://github.com/bjornwallner/DockQ.git
Test sets used in our paper are given in test_sets_pdb/. Given dataset in {db5, sabdab}, Ground truth (bound) structures are in test_sets_pdb/{dataset}_test_random_transformed/complexes/, while unbound structures (i.e., randomly rotated and translated ligands and receptors) are in test_sets_pdb/{dataset}_test_random_transformed/random_transformed/ and you should precisely use those for your predictions.
The best validated model on DB5.5 (db5_best.ckpt) and the best validated model on SAbDab (sabdab_best.ckpt) have been stored in checkpoints. You can inference docked complexes and obtain the same metrics listed in our paper with the following command:
python test.py --dataset DB5 --gpu {GPU} --ckpt ./checkpoints/db5_best.ckptAfter the inference step, the docked complex PDB files are saved in checkpoints/db5_best_results.
To obtain metrics of other methods, please run:
python test.py --model_type {HDock/Multimer/DiffDock-PP/EquiDock} --dataset {DB5/SAbDab} --gpu {GPU}If you only want to generate the docked complex structure without evaluation, use the following command:
python inference.py --dataset {DB5/SabDab} --gpu {GPU} --ckpt {ckpt_path}MIT