ROSE-CUDA is a high-performance, GPU-accelerated fork of the ROSE (Rank Ordering of Super-Enhancers) pipeline originally developed by St. Jude Children's Research Hospital.
This project was entirely developed and validated using Google's Antigravity, leveraging the synergistic power of Opus 4.6 and Gemini 3.1 Pro models. The integration of generative AI and software engineering allowed for the transformation of a complex serial algorithm into a high-performance parallel solution in record time.
- CUDA Acceleration: Custom GPU kernels for BAM read density mapping and interval overlap computation.
- Python Optimization: Replaced legacy
os.system()calls withpysamandNumPyvectorization, achieving up to 100x speedup in core modules. - Parallel Orchestration: Multi-threaded BAM processing and dynamic GPU memory batching for large-scale genomic datasets.
- End-to-End GPU Coverage: All pipeline stages — density mapping, region stitching, and gene mapping — use batch GPU overlap queries.
- O(log N) Closest Gene: Gene-to-enhancer mapping uses binary search (
bisect) instead of the original 50 Mbp linear scan. - Scientific Reproducibility: Includes a COMPAT MODE that ensures bit-for-bit identical results to the original St. Jude implementation.
Create the environment using the provided environment.yml:
conda env create -f environment.yml
conda activate rose-cudaUsage remains fully compatible with the original ROSE:
python bin/ROSE_main.py -g HG38 -i input_peaks.gff -r factor_rankby.bam -o output_dirUse the --no-gpu flag to run in high-performance CPU mode if an NVIDIA GPU is not available.
Developed with Antigravity by Google (Gemini CLI)