PyRHE

PyRHE is a unified and efficient Python framework for genomics heritability estimation. It provides a modular and extensible platform for implementing various genetic architecture estimation models and computation optimizations for large-scale genomic data.

The full documentation is available at the PyRHE Documentation.

Key Features

• High computational efficiency through distributed jackknife subsamples and parallelized genotype I/O and large-scale matrix operations.
• Tensor-based computation with automatic conversion of large matrices to PyTorch tensors, designed to run efficiently on both CPU and CUDA-enabled GPU architectures.
• Memory-efficient streaming support through the StreamingBase class, enabling memory-efficient processing of large-scale genomic data.
• Modular, extensible design with abstract base classes (Base, StreamingBase) that provide interfaces for adding new models.
• Multiple models in one framework, including RHE, RHE-DOM, and GENIE, all sharing common infrastructure.

Installation

pip install pyrhe
# Also install proper version of PyTorch from https://pytorch.org/ 

Example Usage

from pyrhe.models import (
    RHE,
    StreamingRHE,
    GENIE,
    StreamingGENIE,
    RHE_DOM,
    StreamingRHE_DOM,
)

# Standard RHE
rhe_model = RHE(
    geno_file="path/to/genotype",
    annot_file="path/to/annotation",
    pheno_file="path/to/phenotype",
    # other arguments...
)
rhe_results = rhe_model()

# Streaming RHE
streaming_rhe_model = StreamingRHE(
    geno_file="path/to/genotype",
    annot_file="path/to/annotation",
    pheno_file="path/to/phenotype",
    # other arguments...
)
streaming_results = streaming_rhe_model()

Each model (e.g., RHE, GENIE, RHE_DOM and their streaming version) follows the same pattern: initialize with file paths and options, then call the model instance to run estimation and return results.

Run analysis

After installing the package, you can run PyRHE directly from the command line:

python run_rhe.py <command-line arguments>

Alternatively, you may run PyRHE using a newline-separated config file:

python run_rhe.py --config <config file>

Parameters

model: The model to run (e.g., rhe, rhe_dom, genie).
genotype (-g): The path of PLINK BED genotype file
phenotype (-p): The path of phenotype file
covariate (-c): The path of covariate file
annotation (-annot): The path of genotype annotation file.
num_vec (-k): The number of random vectors (10 is recommended). 
num_block (-jn): The number of jackknife blocks (100 is recommended). 
    The higher the number of jackknife blocks, the higher the memory usage.
output (-o): The path of the output file prefix
streaming: Whether to use the streaming version or not
num_workers: The number of workers
seed (-s): The random seed
device: Device to use (cpu or gpu)
      Using CPU already enables great performance. You can further improve performance using GPU
cuda_num: CUDA number of GPU
geno_impute_method: How to impute missing genotype ("binary" (binary imputation) or "mean" (mean imputation))
cov_impute_method: How to impute missing covariate ("ignore" (ignore individuals with missing covariate) or "mean" (mean imputation))
samp_prev: Sample prevalence of binary phenotype (for conversion to liability scale)
pop_prev: Population prevalence of binary phenotype (for conversion to liability scale)
trace (-tr): Save the stochastic trace estimates as trace summary statistics (.trace) with metadata (.MN)
trace_dir: Directory to save the trace estimates

Please refer to the example for a list of configuration files for running RHE, RHE_DOM, and GENIE, and their respective outputs.

Accuracy of Estimation

Runtime Comparison