gpuZoo: Cost-effective estimation of gene regulatory networks using the Graphics Processing Unit

Marouen Ben Guebila, Daniel Morgan, Kimberly Glass, Marieke L. Kuijjer, Dawn L. DeMeo, John Quackenbush

TL;DR

The gpuZoo framework provides GPU-accelerated implementations of popular gene regulatory network (GRN) inference tools like PANDA and LIONESS. By offloading heavy matrix operations to the GPU, the authors achieved up to a 61x speedup in runtime and a 28x reduction in computational costs compared to traditional multi-core CPU approaches.

What Problem Does This Paper Address?

Understanding how genes are regulated—which "master switches" (transcription factors) turn which genes on or off—is central to biology and medicine. Tools like PANDA (Passing Attributes between Networks for Data Assimilation) are powerful for inferring these networks by integrating multiple types of biological data. However, PANDA and its cousins (PUMA, SPIDER, LIONESS) rely on massive, repetitive matrix operations.

As genomic datasets grow to include tens of thousands of genes and thousands of individual samples, the "computational tax" becomes prohibitive. For example, generating sample-specific networks for the GTEx dataset originally took months on a high-performance CPU cluster. This bottleneck prevents researchers from exploring new hypotheses or scaling their studies to the level required for precision medicine.

Methods

The authors developed gpuZoo, a suite of tools that re-implements these algorithms to run on NVIDIA GPUs using CUDA.

  • gpuPANDA: This tool parallelizes the core "Tfunction" (a modified Tanimoto similarity metric) and other matrix operations. Instead of calculating TF-gene interactions sequentially on a CPU, thousands of GPU cores compute these interactions simultaneously.
  • gpuLIONESS: LIONESS is used to estimate networks for individual samples by iteratively "leaving one out" and recalculating the aggregate network. gpuLIONESS parallelizes this process across multiple GPU devices (like the Tesla V100 or P100) using a Message Passing Interface (MPI) hybrid structure.
  • On-line Co-expression: To further optimize LIONESS, the authors introduced an "on-line" derivation of the co-expression matrix. Instead of recomputing the entire matrix from scratch for every sample, they use a mathematical shortcut to update the matrix based on the excluded sample, significantly reducing redundant calculations.

Datasets

The authors benchmarked their tools using three scales of biological data:

  1. Small Network: 652 Transcription Factors (TFs) by 1,000 genes, derived from lymphoblast cell line (LCL) data.
  2. Protein-coding Network: 652 TFs by 27,149 genes (the standard scale for most human biology studies).
  3. Transcript Network: 1,603 TFs by 43,698 transcripts, representing the "maximal capacity" test including alternative splicing variants.

Benchmarks were performed on Amazon Web Services (AWS) using various CPU (c5d, r5a) and GPU (V100, P100, K80) configurations.

Key Insights and Ideas

  • Matrix Symmetry as an Optimizer: The authors optimized memory usage by only storing and computing half of the symmetrical co-expression matrices, a crucial move given that GPU memory (VRAM) is often more limited than system RAM.
  • The Precision Trade-off: While double precision is the gold standard, the paper demonstrates that switching to single precision allows much larger networks to fit into GPU memory with negligible loss in biological accuracy. The "noise" introduced by single precision is far smaller than the natural noise found in gene expression measurements.
  • Cost vs. Speed: A faster instance isn't always more expensive. Because AWS bills by the second, a GPU instance that is 60x faster but only 2x more expensive per hour results in a massive net saving (28x cheaper in this case).
  • The "On-line" Advantage: The mathematical optimization for co-expression becomes increasingly beneficial as the number of samples grows relative to the number of genes, making it a "future-proof" strategy for massive biobanks.

Key Results

  • Massive Speedups: gpuPANDA was up to 61 times faster than memory-optimized CPUs when running on an NVIDIA V100.
  • Drastic Cost Reduction: The GPU implementation was up to 28 times less expensive to run on cloud infrastructure than the CPU equivalent.
  • Scalability: For a population of 127 individuals, gpuLIONESS achieved a 10x to 24x speedup.
  • Memory Efficiency: The GPU implementation showed a 2.6-fold decrease in memory usage compared to the CPU version, thanks to optimized data handling and symmetry exploitation.

Limitations and Future Work

The primary limitation is GPU VRAM. Even with a 32GB Tesla V100, the largest "transcript-level" networks (43k+ rows/columns) could not be loaded in double precision. While single precision offers a workaround, extremely large multi-omic datasets will continue to push the boundaries of current hardware.

The authors suggest that as GPU hardware continues to evolve (with larger memory capacities and faster interconnects), even more complex models—such as those incorporating non-coding RNAs or epigenetic markers across millions of cells—will become routine. gpuZoo is currently available as part of the netZooM (MATLAB) and netZooPy (Python) packages.

Summary updated

Abstract

Gene regulatory network inference allows for the modeling of genome-scale regulatory processes that are altered during development, in disease, and in response to perturbations. Our group has developed a collection of tools to model various regulatory processes, including transcriptional (PANDA, SPIDER) and post-transcriptional (PUMA) gene regulation, as well as gene regulation in individual samples (LIONESS). These methods work by postulating a network structure and then optimizing that structure to be consistent with multiple lines of biological evidence through repeated operations on data matrices. Although our methods are widely used, the corresponding computational complexity, and the associated costs and run times, do limit some applications. To improve the cost/time performance of these algorithms, we developed gpuZoo which implements GPU-accelerated calculations, dramatically improving the performance of these algorithms. The runtime of the gpuZoo implementation in MATLAB and Python is up to 61 times faster and 28 times less expensive than multi-core CPU implementation of the same methods. gpuZoo is available in MATLAB through the netZooM package https://github.com/netZoo/netZooM and in Python through the netZooPy package https://github.com/netZoo/netZooPy.
BibTeX
@inproceedings{guebila2022gpuzoo,
  author = {Marouen Ben Guebila and Daniel Morgan and Kimberly Glass and Marieke L. Kuijjer and Dawn L. DeMeo and John Quackenbush},
  title = {{gpuZoo: Cost-effective estimation of gene regulatory networks using the Graphics Processing Unit}},
  booktitle = {NAR},
  year = {2022},
  doi = {10.1093/nargab/lqac002},
}

Loading related content...

Figure