Genome-scale metabolic models are large and often require repeated linear-programming optimisation, which can strain CPU-based solvers. A preprint on arXiv introduces gpuGEM, an open-source Python library that offloads this optimisation to NVIDIA GPUs via the cuOpt solver. The library adapts its settings to problem size and is designed as a drop-in tool for existing workflows.

In benchmarks across ten models spanning four orders of magnitude in problem size, gpuGEM achieved speed-ups of 1.1x to 94x relative to the faster of HiGHS dual-simplex and interior-point methods for each model. Against Gurobi, it was faster on three of the ten models, with speed-ups of 1.6x to 2.5x. The authors also tested gpuGEM in a candidate-gene prioritisation workflow, where end-to-end query time fell from 41 seconds to 21 seconds.

gpuGEM is released under the MIT licence and available on PyPI. It requires an NVIDIA GPU with Volta architecture or newer, Python 3.11 or later, CUDA 12, and cuOpt 26.6 or later. The paper notes that the tool is validated against established CPU solvers but does not claim to replace them in all cases.