NVIDIA unveils bionemo inference runtime for accelerating protein structure prediction

NVIDIA has introduced BioNeMo Inference Runtime (BioIR), a Python library designed to enhance the performance of structure-prediction models on NVIDIA GPUs. The tool maintains the standard PyTorch workflow while optimizing key operations such as Pairformer and Evoformer stacks, triangle operations, and pairwise attention. BioIR has already been deployed at scale, contributing to the expansion of the AlphaFold Database by generating high-confidence protein-complex structures across 4,777 proteomes. Benchmarking results show BioIR achieves speedups of up to 2.90x in Boltz-2 folding throughput on 8xH100 GPUs, with a rate of 58.5K residues per GPU-hour. NVIDIA’s tests on 1,000 human dimer targets demonstrated significant energy efficiency improvements, estimating a 23 to 43 MWh reduction per million targets compared to a public implementation. The tool is available as an open-source GitHub repository, requiring Python 3.12+, a compatible NVIDIA GPU, and per-chain A3M MSAs.