Skip to content

Latest commit

 

History

History
61 lines (51 loc) · 3.1 KB

README.md

File metadata and controls

61 lines (51 loc) · 3.1 KB
Raw

Guided-PLI

A Transferability-guided Protein-Ligand Interaction Prediction Method
A novel transferability-guided protein-ligand interaction prediction method that effectively fuses multiple modalities while leveraging transferability metrics to guide knowledge transfer during fine-tuning. This method addresses two key challenges in PLI prediction: 1) integrating heterogeneous protein and ligand data modalities, and 2) optimizing the transfer of beneficial pretraining knowledge while avoiding negative transfer.

Directory Structure

├── AttentiveFP/           # GAT model for extracting drug features
├── data/                  # PLI task datasets
├── models/                # The Guided-PLI prediction model
├── args.yaml              # Directories and drug molecule parameters
├── config.py              # Configuration file for parameter settings
├── data_handler.py        # PLI data processing tool
├── main.py                # Main program
├── otfrm.py               # The transfer loss defination based on OTFRM
├── README.md              # Readme file
├── requirements.txt       # Environment dependencies
├── train_test.py          # Engine for training and testing the model
├── utils.py               # Collection of utilites

1. Environment building

python >3.9.17 torch-1.11.0

conda create -n GuidedPLI python==3.9.17
conda activate GuidedPLI
cd Guided-PLI
pip install -r requirements.txt

2. Data Preparation

Place the processed datasets for PDBbind, Kinase, and DUD-E in the data/ directory. Below is a sample entry for each of the curated datasets:

  • PDBbind
PDB-ID seq rdkit_smiles label set
11gs PYTVV...GKQ CC[C@@H](CSC[C@H]...C(=O)c1ccc(OCC(=O)O)c(Cl)c1Cl 5.82 train
  • Kinase
PDB-ID seq rdkit_smiles label set
Q14012 MLGA...HQL N#Cc1ccc(NC(=O)Nc2ccnc3cc(C(F)(F)F)ccc23)nc1 0.0 train
  • DUD-E
PDB-ID seq rdkit_smiles label set
andr FLNV...HTQ O=c1cc(-c2ccccc2)[nH]c2cc(-c3ccc(I)cc3)nn12 0 train

3. Pretrained Embedding Generation

To guide the finetuning with transferability from pretraining of MASSA, the embeddings of pretraining datasets from MASSA need to be generated first. More details please refer to MASSA repository.

4. Fine-tuning on PLI Tasks

Run main.py to perform fine-tuning from pre-trained PLMs to downstream PLI prediction. The following example demonstrates the command to fine-tune on the PDBBind task with guidance from transferability:

python main.py --task=PDBBind --SEED=42 --guide

For more input parameter settings, please refer to config.py.

Acknowledgement

The SOFTWARE will be used for teaching or not-for-profit research purposes only. Permission is required for any commercial use of the Software.