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

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.

Graphical Abstract

Gene regulatory network (GRN) inference for large studies using PANDA and LIONESS can provide thousands of regulatory network models that can provide insight into biological systems; GPU-optimized code can greatly accelerate the process of network inference and substantially reduce costs.

Figure 1.

Runtime and cost performance of gpuPANDA in the small network. (A). Runtime (first row) and cost (second row) fold change between CPU1, CPU2, GPU2, and GPU3. The results are reported as CPU/GPU fold change, therefore, a fold change larger than 1 indicates a decrease of cost or runtime using GPU. Conversely, a fold change less than 1 indicates an increase in cost or runtime using GPU. (B). Rate of cost fold change as an effect of runtime fold change in the small network in single and double precision. The blue area represents the case when CPU/GPU fold change is less than 1 indicating an increase in cost and/or runtime of GPU computation over CPU.

Figure 2.

Runtime and cost performance of GPUs and CPUs on the protein coding-genes model. (A). Fold change of runtime as a function of cost between CPU1 and GPU2 and CPU2 and GPU2 in single precision and for three values of learning rate (α). A CPU/GPU fold change larger than 1 indicates a decrease in cost or runtime using GPU. Fold change smaller than 1 indicates an increase in runtime or cost using GPU. (B). Effect of runtime fold change on cost fold change between CPU1 and GPU2 (top panel) and CPU2 and GPU2 (bottom panel). The blue area indicates the cases where using GPU is slower and/or more expensive than CPU (fold change < 1).

Figure 3.

GPU performance on transcript model and memory benchmark. (A). Memory usage of GPU implementation in comparison to CPU implementation. (B). Tested network precision for the tested hardware using the small network, protein-coding genes network, and transcript regulatory gene network. (C). Runtime and D-cost of running transcript model on GPU1, CPU1, and CPU2 in single precision. Average coefficient of variation across learning rates in runtime is 0.4% for CPU1, 0.5% for CPU2, and 0.2% for CPU3. *Single precision computation on GPU1 converges with Tfunction only.