InSiteDTA

A complex-free deep learning model for protein-ligand binding affinity prediction with intrinsic binding site detection.

Key Features:

• No molecular docking required
• Robust performance regardless of binding site determination method
• Robust performance on imperfect structural inputs

Installation

1. Clone repository

git clone https://github.com/KU-MedAI/InSiteDTA.git
cd InSiteDTA

2. Create conda environment

conda env create -f environment.yml
conda activate insite

3. Install P2Rank (Optional, Recommended) — Krivák & Hoksza, 2018

mkdir src/p2rank && cd src/p2rank
wget https://github.com/rdk/p2rank/releases/download/2.5.1/p2rank_2.5.1.tar.gz
tar -xzf p2rank_2.5.1.tar.gz -C ./ --strip-components=1

Why P2Rank? InSiteDTA internally predicts the binding site and uses it as a feature for affinity prediction, so P2Rank is not strictly required. However, providing a P2Rank-predicted pocket helps guide the voxelization step so that the sampled protein voxel is more likely to include the true binding site. This can enable more sophisticated prediction, especially when inferencing with large proteins.

Our tested environment:

• Python: 3.9.19
• PyTorch: 2.5.1
• PyTorch Geometric: 2.6.1
• CUDA: 11.8
• P2Rank: 2.5.1

Quick Start Example

Without pocket guidance (unguided voxelization):

python 01-inference.py \
    --pdb_path ./src/data/samples/4gkm/4gkm_protein.pdb \
    --smiles "Cc1ccc(c(c1)C(=O)[O-])Nc1ccccc1C(=O)[O-]"

With P2Rank guidance (guided voxelization, recommended):

python 01-inference.py \
    --pdb_path ./src/data/samples/4gkm/4gkm_protein.pdb \
    --smiles "Cc1ccc(c(c1)C(=O)[O-])Nc1ccccc1C(=O)[O-]" \
    --use_p2rank

Training With Your Own Data

Step 1: Prepare Data Structure

Organize your data in nested structure (PDBbind format):

raw_data/
├── {pdb_id}/
│   ├── {pdb_id}_protein.pdb
│   └── {pdb_id}_pocket.pdb
...

Prepare SMILES CSV file (smiles.csv):

PDB_ID,Canonical SMILES
1abc,CCO
1def,c1ccccc1

For affinity prediction, prepare affinity index JSON (affinity.json):

{"1abc": 5.2, "1def": 7.8}

Note: If you only want to train binding site prediction, omit --index_file argument in preprocessing.

Step 2: Preprocess

python 02-preprocess.py \
    --raw_dir ./raw_data \
    --save_dir ./preprocessed \
    --smiles_csv ./smiles.csv \
    --index_file ./affinity.json \
    --test_key_file ./test_keys.txt \
    --voxel_size 2 \
    --n_voxels 32 \
    --device 0

This generates preprocessed data and data_config_*.json in ./preprocessed/.

Step 3: Train

python 03-train.py \
    --data_config ./preprocessed/data_config_*.json \
    --save_dir ./checkpoints \
    --device 0 \
    --epochs 300 \
    --batch_size 48

Trained model will be saved as ./checkpoints/{timestamp}_{data_config_name}.pt.

Evaluate Your Trained Model

python 04-evaluate.py \
    --ckpt ./checkpoints/{experiment_name}.pt \
    --result_file ./checkpoints/{experiment_name}_results.json \
    --save_dir ./evaluation \
    --device 0

The script will:

1. Load the test split defined in the training result file
2. Run inference on the test set
3. Report performance metrics (PCC, RMSE, MAE, DCC, DVO)
4. Save detailed results to {save_dir}/{experiment_name}_test_results.csv

Reproduce Paper Results

Run evaluation on three benchmark datasets:

# Evaluate on Coreset_crystal
python 05-reproduce.py --data crystal --batch_size 64 --device 0

# Evaluate on Coreset_redocked  
python 05-reproduce.py --data redocked --batch_size 64 --device 0

# Evaluate on Coreset_p2rank
python 05-reproduce.py --data p2rank --batch_size 64 --device 0

The script will:

1. Prepare ligand features from SMILES
2. Voxelize protein structures
3. Evaluate with three trained models
4. Report performance metrics (PCC, RMSE, MAE)

Output

Inference (01-inference.py):

• Predicted binding affinity in pK scale (higher values = stronger binding)

Training (03-train.py):

• Model checkpoint: {save_dir}/{timestamp}_{data_config_name}.pt
• Training results: {save_dir}/{timestamp}_{data_config_name}_results.json

Evaluate (04-evaluate.py):

• Evaluation results CSV: {save_dir}/{experiment_name}_test_results.csv

Reproduce (05-reproduce.py):

• Performance metrics (mean ± std across 3 models): PCC, RMSE, MAE

Data

$Coreset_{crystal}$

• Standard benchmark dataset from PDBbind

$Coreset_{redocked}$

• Coreset with redocked ligand in the native pocket

$Coreset_{p2rank}$

• Coreset with redocked ligand in the p2rank predicted pocket

Citation

TBD