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

Abstract

Gene regulatory network inference allows for the study of transcriptional control to identify the alteration of cellular processes in human diseases. Our group has developed several tools to model a variety of regulatory processes, including transcriptional (PANDA, SPIDER) and post-transcriptional (PUMA) gene regulation, and gene regulation in individual samples (LIONESS). These methods work by performing repeated operations on data matrices in order to integrate information across multiple lines of biological evidence. This limits their use for large-scale genomic studies due to the associated high computational burden. To address this limitation, we developed gpuZoo, which includes GPU-accelerated implementations 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 the multi-core CPU implementation of the same methods. gpuZoo takes advantage of the modern multi-GPU device architecture to build a population of sample-specific gene regulatory networks with similar runtime and cost improvements by combining GPU acceleration with an efficient on-line derivation. Taken together, gpuZoo allows parallel and on-line gene regulatory network inference in large-scale genomic studies with cost-effective performance.

Introduction

Understanding the complex landscape of transcriptional control—specifically how transcription factors (TFs) act as "master switches" to regulate gene expression—is fundamental to modern biology and precision medicine. Tools such as PANDA (Passing Attributes between Networks for Data Assimilation) and its derivatives (PUMA, SPIDER) have proven invaluable for inferring these networks by integrating diverse biological data types. However, these methods rely on massive, repetitive matrix operations that create a significant "computational tax."

As genomic datasets expand to include tens of thousands of genes and thousands of individual samples, traditional CPU-based processing becomes a bottleneck. For instance, generating sample-specific networks for large-scale repositories like the Genotype-Tissue Expression (GTEx) project can take months on high-performance CPU clusters. To overcome this, we developed gpuZoo, a framework designed to offload these heavy computations to Graphics Processing Units (GPUs), drastically reducing both time and financial costs.

The gpuZoo Framework

gpuZoo provides GPU-accelerated implementations of popular gene regulatory network (GRN) inference tools, re-engineered to run on NVIDIA GPUs using CUDA. The framework focuses on two primary areas: aggregate network inference and sample-specific network estimation.

gpuPANDA

The core of the PANDA algorithm involves calculating a modified Tanimoto similarity metric (the "Tfunction") across large matrices. gpuPANDA parallelizes these operations, allowing thousands of GPU cores to compute TF-gene interactions simultaneously rather than sequentially. This approach is particularly effective for the dense matrix multiplications that define the message-passing process in PANDA.

gpuLIONESS and On-line Optimization

LIONESS is used to estimate networks for individual samples by iteratively "leaving one out" and recalculating the aggregate network. To optimize this computationally intensive process, gpuZoo employs two strategies:

  1. Multi-GPU Parallelization: gpuLIONESS distributes the workload across multiple GPU devices (such as the Tesla V100 or P100) using a hybrid Message Passing Interface (MPI) structure.
  2. On-line Co-expression Derivation: Instead of recomputing the entire co-expression matrix from scratch for every sample, we introduced a mathematical shortcut to update the matrix based on the excluded sample. This "on-line" approach significantly reduces redundant calculations as the sample size grows.

Performance Benchmarking

We evaluated gpuZoo using three scales of biological data, ranging from small networks (652 TFs by 1,000 genes) to massive transcript-level networks (1,603 TFs by 43,698 transcripts). Benchmarks were conducted on Amazon Web Services (AWS) to compare various CPU and GPU configurations.

Speed and Cost Efficiency

The results demonstrate a transformative shift in performance. gpuPANDA achieved up to a 61x speedup on an NVIDIA V100 compared to memory-optimized multi-core CPUs. Because cloud providers bill by the second, this speedup directly translates to cost savings; the GPU implementation was found to be up to 28 times less expensive than the CPU equivalent.

Memory and Precision

GPU memory (VRAM) is often more constrained than system RAM. To address this, gpuZoo utilizes several optimizations:

  • Symmetry Exploitation: By only storing and computing half of the symmetrical co-expression matrices, we achieved a 2.6-fold decrease in memory usage.
  • Precision Trade-offs: While double precision is standard, we found that switching to single precision allows larger networks to fit into VRAM with negligible loss in biological accuracy. The numerical noise introduced by single precision is significantly lower than the inherent noise in gene expression measurements.

Conclusion

gpuZoo represents a significant step forward in making large-scale genomic studies accessible and affordable. By leveraging the parallel processing power of modern GPUs and optimizing underlying mathematical operations, we have reduced the time required for complex network inference from months to days. gpuZoo is currently available through the netZooM (MATLAB) and netZooPy (Python) packages, providing a scalable solution for the next generation of multi-omic research.

Summary updated

Abstract

ABSTRACT Gene regulatory network inference allows for the study of transcriptional control to identify the alteration of cellular processes in human diseases. Our group has developed several tools to model a variety of regulatory processes, including transcriptional (PANDA, SPIDER) and post-transcriptional (PUMA) gene regulation, and gene regulation in individual samples (LIONESS). These methods work by performing repeated operations on data matrices in order to integrate information across multiple lines of biological evidence. This limits their use for large-scale genomic studies due to the associated high computational burden. To address this limitation, we developed gpuZoo, which includes GPU-accelerated implementations 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 the multi-core CPU implementation of the same methods. gpuZoo takes advantage of the modern multi-GPU device architecture to build a population of sample-specific gene regulatory networks with similar runtime and cost improvements by combining GPU acceleration with an efficient on-line derivation. Taken together, gpuZoo allows parallel and on-line gene regulatory network inference in large-scale genomic studies with cost-effective performance. 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},
}
Figure