FedGS: Genomic Selection with Federated Learning

​The program 'fedgs' can be used to perform genomic selection with our proposed federated learning algorithm. In this program, phenotypes of your own data can be predicted using our algorithms. The effectiveness of FedGS has been demonstrated on a wide range of datasets, both simulated and real-world.

Installation

Docker Installtion

Go to docker website and download the version satisfying your machine.

FedGS Docker image installation and quickly start

docker pull linjie7674/fedgs
docker run -it linjie7674/fedgs python3 dockerMain.py kfold --isFed False --genotype test 

Note: The test data is solely utilized for program correction testing purposes and does not hold any significance. If the aforementioned commands are executed accurately, your installation has been successful.

Prepare Data

This project used many datasets.It contains public data and private data.

• test data

  test data is only used to test the program correction.

• your own data

  You can use your own data, which must be orginized by the follow structure.

  - data
      - dataset_name your dataset name, which you can define it by yourself
          - datas.h5 this file includes your genotype data, where the first column should represent id 
          number, and the others should be genotype data
          - labels_*.h5 this file includes your label data, where the first column should be the id number,
          and the others should be the labels corrsponding to id number. The `*` repersents that you can
          place different labels in your directory and you can define every labels data's name.
  

Run

Train using 10-fold cross validation

docker run -it -v {host directory of your own data}:/fedgs/data linjie7674/fedgs python3 dockerMain.py kfold --isFed True --genotype {local dataset name} --label_suf {local label suffix} --fed_genotyp {others dataset name} --fed_label_suf {others label suffix}

Note: Supposing that user's private dataset is located in directory /home/test , the change the words above ({host directory of your own data}) to use's directory /home/test.

• isFed: train solely or federated learning
• genotype: directory name of local dataset directory
• label_suf: label suffix of local dataset
• fed_genotype: external dataset directory name, which will be - federated learning with your local data.
• fed_label_suf: label suffix of external dataset

Train Purely

Train without 10-fold cross validation.

docker run -it -v {host directory of your own data}:/fedgs/data linjie7674/fedgs python3 dockerMain.py train --isFed True --dataset_name {data1} --fed_dataset_name {data2} --batch_size 28 --lr 0.001 --study_name fedgs

The data used for pure training should be organized in this way.

- data
    - dataset_name
        - {dataset_name}_train_datas.h5
        - {dataset_name}_valid_datas.h5
        - {dataset_name}_test_datas.h5

Test

Test is to get model accuracy.

docker run -it -v {host directory of your own data}:/fedgs/data linjie7674/fedgs python3 dockerMain.py test --dataset_name {directory of data} --model_path {path of trained model}

Inference

Inference is to obtain trait by genotype.

docker run -it -v {host directory of your own data}:/fedgs/data linjie7674/fedgs python3 dockerMain.py infer --dataset_name {directory of data} --model_path {path of trained model}

Run on GPU

docker run -it --gpus all linjie7674/fedgs python3 dockerMain.py ...

Get more information about the command

docker run -it linjie7674/fedgs python3 dockerMain.py --help