This GitHub repository contains the dataset and relevant code for the paper:
SIU: A Million-Scale Structural Small Molecule-Protein Interaction Dataset for Unbiased Bioactivity Prediction
dataset is available at huggingface dataset repo
Please download and unzip all files
it should contains:
The dataset used to train gnn and cnn-3d model.
The dataset used to train Uni-Mol or ProFSA model for the SIU 0.6 version. It contains train.lmdb, valid.lmdb and test.lmdb.
The dataset used to train Uni-Mol or ProFSA model for the SIU 0.9 version. It contains train.lmdb, valid.lmdb and test.lmdb.
Pretrained weights for Uni-Mol and ProFSA.
profsa.pt: weights for pretrained ProFSA model
mol_pre_no_h_220816.pt: weights for pretrained Uni-Mol molecular Encoder
pocket_pre_220816.pt: weights for pretrained Uni-Mol pocketr Encoder
Complete dataset file in pickle format
each key is an UniProt ID
corresponding value is a list of dictionaries. Each dictionary is a data point and has following keys:
| Key | Description |
|---|---|
| atoms | atom types in ligand |
| coordinates | list of different conformations of the ligand |
| pocket_atoms | atom types in pocket |
| pocket_coordinates | atom positions of the pocket |
| source_data | UniProt ID and PDB ID information |
| label | dictionary for assay types and assay values |
| ik | InChI key of the ligand |
| smi | SMILES notation of the ligand |
All training and testing data are in lmdb format, and have the same keys as shown above.
Note that for single task learning, the label is a float value instead of a dictionary.
In read_data.py, we provide a script to read from lmdb files and pickle files.
Follow the environment setting in Uni-Mol and Atom3D
Or use siu.yaml
cd ./atom3d/examples/lbafor cnn3d
cd cnn3d for gnn
cd gnn start training
python train.py
Note that the data path in train.py should be changed to atom3d_data/split_60 or atom3d_data/split_90
All the parameters are in train.py. We train the model with one NVIDIA A100 GPU.
cd ./unimol_train_code
bash binding_affinity_unimol.sh or binding_affinity_profsa.sh
use the pretrained weights in pretrain_weights.zip
Note that the data_path should in the bash file should be pointed to dir split_60 or split_90 (0.6 version and 0.9 version respectively)
If you want to to Multi Task Learning, please let --num-heads equals to 5, else set it to 1 and point to the correct directory in split_60 or split_90 (ic50, ec50, ki, kd)
All the parameters are in bash scripts. We train the model with 4 NVIDIA A100 GPU.