OmniESI

The official code repository of "OmniESI: A unified framework for enzyme-substrate interaction prediction with progressive conditional deep learning"

News

⭐Aug 26, 2025: We have released the code, model weights and dataset for active site prediction! Check OmniESI_site.zip for more details.😊

⭐Jul 10, 2025: The source code for OmniESI is released!

⭐Jun 25, 2025: More downstream applications coming in one week. Stay tuned!😊

Overview

Installation

Create a new environment for OmniESI:

conda create -n OmniESI python=3.8 -y
conda activate OmniESI

Installation for pytorch 1.12.1:

pip install torch==1.12.1+cu113 torchvision==0.13.1+cu113 torchaudio==0.12.1 --extra-index-url https://download.pytorch.org/whl/cu113

Installation for other dependencies:

pip install dgl-cu113 -f https://data.dgl.ai/wheels/repo.html
chmod +x scripts/setup_env.sh
bash ./scripts/setup_env.sh

Data preparation

Download ESI datasets and model weights at:

Download link

Prepare ESI datasets

Make sure all the ESI datasets are stored under:

/.../OmniESI/datasets/*

Prepare model weights of OmniESI

We have released all the weights for the main ESI tasks. Make sure all the weights are stored under:

/.../OmniESI/results/*

Obtain ESM-2(650M) embeddings

Generate embeddings for all ESI datasets. First specify the storage path for the ESM embeddings: [DATA_PATH]. The following script will save the embeddings to [DATA_PATH]/dataset_name/esm/. By default, this will retrieve the ESM features for all datasets in the OmniESI paper. We recommend running it on a GPU with at least 24GB memory

cd ./scripts
python embedding.py --feat_dir [DATA_PATH]
cd ..

!!NOTE: We highly recommand employ the absolute path /.../OmniESI/datasets_embeddings for [DATA_PATH]. This will ensure the quick usage of subsequent code and the reproducibility of results.

Quick reproduction

Once the ESM-2 embeddings and all model weights are prepared, you can quickly reproduce the main experimental results mentioned in the paper

Inference reproduction

chmod +x ./scripts/reproduce_inference.sh
./scripts/reproduce_inference.sh

On a single V100 GPU, it takes approximately one hour to obtain all the results.

Training reproduction

chmod +x ./scripts/reproduce_training.sh
./scripts/reproduce_training.sh

On 4 V100 GPU, it takes approximately 2-3 days to obtain all the results.

Custom data inference

You can specify any model weights to perform inference on custom data. First, you need to organize the enzyme-substrate pairs (enzyme sequences and substrate SMILES) you want to predict into a CSV file containing two columns: "SMILES" and "Protein".

Protein,SMILES,
seq1,smile1,
seq2,smile2,
...

Then, select the model weights corresponding to the specific task for inference. For example, use the weights from the ESP dataset for pair prediction.

CUDA_VISIBLE_DEVICES=0 python inference.py \
                 --model configs/model/OmniESI.yaml \
                 --csv_data [INPUT CSV] \
                 --weight ./results/esp/OmniESI/best_model_epoch.pth \
                 --task binary

The prediction csv results will be saved in the OmniESI/ directory.