Meta-Learning on Protein Mutation Sequence

Welcome to the Meta-learning for Protein Mutate Property Difference Prediction repository! This repository corresponds to the meta-learning approach on Protein Mutation property predict project.

Getting Started

1. Clone the Repository:

   git clone https://github.com/kuku0202/maml.git
   cd maml
   

2. Create virtual environment to run this project:

   conda create -n maml python=3.9
   conda activate maml
   

3. Install Required Packages:

Ensure you have Python and pip installed. Then, install the necessary packages:

pip install -r requirements.txt

4. Download the model:

To download the model, either:

git lfs install
git clone https://huggingface.co/yuesu4/Protein_Mutation_ProtBert_MAML

or:

from huggingface_hub import snapshot_download

snapshot_download(repo_id="yuesu4/Protein_Mutation_ProtBert_MAML", local_dir="./")

Note: Make sure to install huggingface_hub if using option 2:

pip install huggingface_hub

The model from https://huggingface.co/yuesu4/Protein_Mutation_ProtBert_MAML should be saved in the saved_models directory for use with the pipeline.

Usage

Run the full pipeline on all datasets:

python main.py --initial_task preprocess_data/combined_ddG_all.csv 
    --meta_learning_tasks preprocess_data/binding_affinity_train.csv preprocess_data/fireprot_dTm_train.csv preprocess_data/soluprotmut_solubility_train.csv 
    --test_tasks preprocess_data/binding_affinity_test.csv preprocess_data/fireprot_dTm_test.csv preprocess_data/soluprotmut_solubility_test.csv
    --finetune_train_tasks preprocess_data/binding_affinity_train.csv preprocess_data/fireprot_dTm_train.csv preprocess_data/soluprotmut_solubility_train.csv 
    --run_maml 
    --maml_epochs 50 
    --finetune_epochs 10
    --save_dir ./results

If you have downloaded the pretrained model from huggingface, you can skip the pretraining steps by adding:

python main.py --skip_pretrain --pretrained_model_path your_path_to_store_model_path
......

If you do not want to include target training part as tasks, there is an systematic approach to evaluate every task rather than run it one by one:

  python main.py --initial_task preprocess_data/combined_ddG_all.csv 
    --test_tasks preprocess_data/binding_affinity_test.csv preprocess_data/fireprot_dTm_test.csv preprocess_data/soluprotmut_solubility_test.csv 
    --maml_epochs 50 
    --finetune_epochs 10 
    --save_dir ./results