ZIBRA

ZIBRA is a high-performance computing framework for analyzing gene-gene associations in large-scale single-cell RNA sequencing (scRNA-seq) data using the Bivariate Zero-Inflated Negative Binomial (BZINB) model.

Prerequisites

1. Python Environment

conda create -n bzinb python=3.9
conda activate bzinb
pip install numpy pandas scanpy scipy tqdm statsmodels

2. C++ & HPC Libraries

• OpenMPI or MPICH
• NLopt
• Eigen 3

Quick Start

Step 1: Compile the C++ MPI Engine

Ensure paths to eigen3 and nlopt match your system configuration in the Makefile, then compile:

make

Step 2: Data Preprocessing (Python)

Pack your .h5ad single-cell data into binary chunks:

python scripts/prepare_data.py

Step 3: MPI Computation (C++)

Launch the MPI engine using all available cores (e.g., 124 cores):

mkdir -p output_chunks
mpirun -n 124 ./zibra_mpi_engine

Step 4: Merge Results (Python)

Merge the distributed output chunks and map the numerical IDs back to actual Gene Names:

python scripts/merge_results.py

Output: ZIBRA_Final_Results.csv

Advanced Customization

• Gene Filtering: Modify step1_univariate_filter or the np.mean threshold in scripts/prepare_data.py.

• Significance Threshold: The decision tree threshold (alpha = 1e-8) can be modified directly in src/zibra_mpi.cpp.