A rapid and reference-free imputation method for low-cost genotyping platforms

This repository contains a Python implementation of GRUD model, which is a genotype imputation based on deep learning algorithms. In specific, GRUD is composed of two components: Generator and Discriminator. The generator model undertakes a mission to create tokens, and the discriminator tries to verify tokens created by the generator model. In the current study, we assumed that unobserved variants are tokens in natural languages which would be predicted from a known paragraph of observed variants. In addition to, our approach resolves the reference panel privacy problem while improving both the accuracy and running time.

If you have any feature requests or questions, feel free to leave them as GitHub issues!

REQUIREMENT

• Python = 3.8.8
• Python packages:
  • NumPy
  • Scikit-learn
  • PyTorch

pip install -r requirements.txt

EXAMPLE USAGE

1. Dataset

In this repository, the example dataset that is used to train and evaluate is the 1000 Genome Project (1KGP) phase 3 integrated dataset for chromosome22.

Download: https://drive.google.com/drive/folders/17RIdXWoIKJsxjzfYmSmvW19kaQHxoafC?usp=sharing

• Type of input data:
  • Phased genotype with HAP/LEGEND format ( For details, please see https://github.com/kanamekojima/rnnimp)
  • Weighted of trained model
• Type of output data:
  • Genotype imputation results in Oxford GEN format
  • VCF format

2. Evaluation/ Inference

For example, GRUD model imputes genotypes for small regions (1-10) separately.

In evaluation, the root directory should contain ground truth files while inference doesn't need this.

python eval.py  --root-dir <path of LEGEND/HAP files> \
                --model-config-dir <path of config model files> \
                --model-dir <path of weighted model files> \
                --result-gen-dir <path of output files> \
                --sample <path of sample name file> \
                --regions 1-10 \
                --best-model

3. Training

We provides small regions of chromosome 22 for training 1KGP data.

python train.py --root-dir <path of LEGEND/HAP files> \
                --model-type dis \
                --model-config-dir <path of config model files> \
                --batch-size 128 \
                --epochs 100 \
                --chr chr22 \
                --lr 0.001 \
                --output-model-dir <output path of weighted model files> \
                --early-stopping \
                --gpu 0 \
                --region 1-10

5. Arguments

Args	Default	Description
--model-type STR	dis	Type of model
--root-dir STR	None	Data folder
--model-config-dir STR	None	Config model folder
--gpu STR	None	GPU's Id
--batch-size INT	128	Type of model
--regions STR	1	Range of regions
--chr STR	chr22	Chromosome
--model-dir STR	model/weights	weight model folder
--result-gen-dir STR	results/	result folder
--dataset STR	None	Custom dataset
--lr FLOAT	1e-4	learning rate
--output-model-dir STR	None	Custom dataset
--early-stopping BOOL	False	Early stopping
--best-model BOOL	False	Get best model to test
--sample STR	None	path to sample name file
--thread INT	1	parallel imputation in inference

TODO

• Training model source code
• Evaluation model source code
• Inference model source code (updated on 07/01/2024)
• Parallel in inference (updated on 07/02/2024)
• VCF to HAP/LEGEND format source code

CITATION

If you find GRUD or any of the scripts in this repository useful for your research, please cite:

Chi Duong, V., Minh Vu, G., Khac Nguyen, T., Tran The Nguyen, H., Luong Pham, T., S. Vo, N., & Hong Hoang, T. (2023). A rapid and reference-free imputation method for low-cost genotyping platforms. Scientific Reports, 13(1), 23083. https://doi.org/10.1038/s41598-023-50086-4

LICENSE

The scripts in this repository are available under the MIT License. For more details, see the LICENSE file.

CONTACT

Developer: Duong Chi Vinh, GeneStory

E-mail: vinh.duong [AT] genestory [DOT] ai