PhosLLPS

PhosLLPS, a Transformer-based GNN prediction medthod (AUC=0.9116) to identify functional phosphorylation sites that regulate liquid-liquid phase separation (LLPS)

• The complete protein sequence (with a length of L amino acids) is input into the pre-trained protein language model ESM2 to obtain L x 1280-dimensional features. Then, a 15 x 1280-dimensional feature set is derived as node features (where 15 represents the 7 amino acids upstream and downstream of the phosphorylation sites). A 15 x 15 full adjacency matrix is used as the graph's edge features. Next, the graph is processed through three TransformerConv layers, each employing a self-attention mechanism. Finally, the processed features are input into a multi-layer perceptron (MLP) to generate the final prediction results.

Contact

<1>haifengchen@sjtu.edu.cn <2>hongxk@fzu.edu.cn

Set up environment

1. Copy Python library

• git clone https://github.com/HFChenLab/PhosLLPS.git
• cd PhosLLPS

2. Set up a Conda environment

• conda env create -f environment.yml
• conda activate PLM2

3. install esm software

• pip install fair-esm

4. install biopython software

• conda install -c conda-forge biopython

Obtain embedding features of pretrained protein language model ESM2

• cd esm2
• prepare P00533.fasta
• prepare esm2_batch.csv
• python esm2.py

Extract key features from above embedding features

• cd dataset
• prepare dataset.csv, including uniprot and site
• python generate_dataset.py

Predict prob and pred of each site of proteins

• cd predict
• python predict.py