EquiSite: Multi-Scale Equivariant Graph Learning for Robust Nucleic Acid Binding Site Prediction

This is the official PyTorch implementation of EquiSite.

EquiSite is an SE(3)-equivariant geometric graph neural network for accurate protein-DNA and protein-RNA binding site prediction. EquiSite explicitly models multi-scale protein geometry (including atomic side-chain orientations) and eliminates the need for computationally expensive evolutionary profiles (MSAs).

EquiSite achieves state-of-the-art performance on both experimental structures and AlphaFold2-predicted models, and supports structure-guided molecular docking (e.g., HADDOCK3)

Graphical Abstract

---

Installation

We recommend using Anaconda to manage the environment.

conda env create -f conda_env.yml
conda activate equisite

Core dependencies include:

• Python >= 3.9
• PyTorch
• PyTorch Geometric
• ESM (facebookresearch/esm)
• h5py
• numpy
• scikit-learn

---

Training

Example training command:

python train.py \
    --dataset RNA_Fix \
    --dataset_path dataset/ \
    --epochs 200 \
    --batch_size 4 \
    --eval_batch_size 2 \
    --hidden_channels 128 \
    --num_blocks 4 \
    --cutoff 11.5 \
    --level allatom+esm

Training logs and checkpoints will be saved automatically under:

./saves/

The best model (based on validation ROC-AUC) is saved as:

best.pt

---

Inference (PDB → Binding Probabilities)

EquiSite supports end-to-end inference from raw PDB files.

Example:

python infer.py \
    --model_path path/to/best.pt \
    --pdb_dir path/to/pdb_folder \
    --out_dir output_directory \
    --device 0

For each input structure:

• A per-residue binding probability file (.out) is generated

---

Reproducing Manuscript Results

To reproduce the main results:

1. Download pretrained checkpoints from Zenodo.
2. Place the checkpoint file in a local directory.
3. Run evaluation using infer.py or the evaluation logic in train.py.
4. Compare generated outputs with archived result files.

All evaluation outputs reported in the manuscript are archived in the Zenodo release.

---

Data

Datasets used in this study are derived from previously published benchmark collections.
Please refer to the manuscript for detailed dataset construction and preprocessing procedures.

This repository contains data processing utilities but does not redistribute raw third-party datasets.

---

📄 License

This project is released under the MIT License.

---