NVIDIA's BioIR cuts protein folding energy use by 43 MWh per million targets

The bottleneck in structural biology has shifted. It is no longer about whether an AI model can accurately predict a single protein’s shape, but how quickly it can process thousands of independent targets. NVIDIA’s new BioNeMo Inference Runtime (BioIR) addresses this queue management problem head-on. This Python library accelerates supported structure-prediction models on NVIDIA GPUs without disrupting standard PyTorch workflows. The results are already visible at scale. BioIR powered the recent expansion of the AlphaFold Database, generating structures for 31 million candidate complexes across 4,777 proteomes.

What makes BioIR distinct is its lack of friction. Developers do not need to compile CUDA source, manage CMake files, or export separate engine artifacts. Models remain ordinary torch.nn.Module objects. The library optimizes specific operations like Pairformer stacks and triangle operations through custom kernels and CUDA Graph capture. It offers two usage paths: an end-to-end processor that handles parsing to PDB writing, or direct PyTorch integration for custom code. While it supports ligand structure prediction, users must supply their own MSAs and templates, as BioIR does not run HHsearch internally.

The performance gains are substantial. In a benchmark of 1,000 human dimer targets on eight H100 GPUs, BioIR-accelerated Boltz-2 delivered 58.5K successfully folded residues per GPU-hour. The public torch-compiled implementation managed only 20.2K and ran out of memory on 29 targets. This represents a 2.90x improvement in residue-normalized throughput. Beyond speed, the energy implications are significant. Extrapolating these results to one million targets suggests BioIR could save between 23 and 43 MWh compared to standard open-source implementations. That is a material reduction in operational cost for large-scale campaigns.

This shift toward optimized inference runtimes signals a maturing market. We are moving past the era of proof-of-concept models into production-grade infrastructure where throughput and energy efficiency dictate viability. By removing the need for complex build steps, NVIDIA lowers the barrier for labs to adopt high-performance folding pipelines. However, the reliance on pre-staged MSAs means the upstream data preparation burden remains. As context-parallel folding arrives in future updates, the ability to split single massive targets across GPUs will likely become the next critical frontier.