NVIDIA has released a specialized software engine that accelerates protein-folding predictions nearly threefold, addressing a critical bottleneck in structural biology research. The BioNeMo Inference Runtime (BioIR) is designed to handle the massive computational demands of modern protein-structure prediction at scale, moving beyond single-target runs to process entire proteomes efficiently.

What Changed in Protein-Folding Workflows?

For years, the challenge in protein folding was whether an artificial intelligence model could accurately predict a protein's three-dimensional shape. That problem is largely solved. Today, the real bottleneck is speed and scale. Researchers now need to fold thousands or millions of proteins in parallel, and the computational pipeline involves multiple stages: parsing input data, generating sequence alignments, running GPU inference, and writing results to files. BioIR optimizes all of these steps simultaneously.

The software has already proven itself in production. NVIDIA used BioIR to power a major expansion of the AlphaFold Database, generating protein-complex structures across 4,777 proteomes and producing approximately 31 million candidate complexes, of which 1.81 million were released as high-confidence predictions.

How Does BioIR Achieve These Speed Gains?

• Custom Kernel Selection: The engine chooses optimized computational kernels tailored to specific model configurations, GPU hardware, data types, and tensor shapes, rather than relying on generic inference libraries.
• Module-Level Optimization: BioIR uses CUDA Graph capture to reduce GPU launch overhead, allowing compatible modules to execute with minimal delays between operations.
• Pipeline Scaling with Ray: The system distributes independent protein targets across multiple GPUs on a single node, while overlapping CPU-based tasks like parsing and feature generation with GPU folding to eliminate idle time.

In a head-to-head benchmark, NVIDIA tested BioIR-accelerated Boltz-2 (a protein-folding model) against a standard open-source PyTorch implementation on eight NVIDIA H100 GPUs. Both used identical datasets of 1,000 human protein dimers with combined sequence lengths under 2,800 residues. BioIR completed all 1,000 targets and delivered 58.5K successfully folded residues per allocated GPU-hour, while the public implementation achieved only 20.2K residues per GPU-hour and ran out of memory on 29 targets.

That 2.90x improvement in throughput translates to significant energy savings. Extrapolating the benchmark to one million comparable targets, BioIR would require approximately 11 megawatt-hours of energy versus 35 megawatt-hours for the standard approach, using eight-GPU thermal design power equivalents. The actual savings could be even larger when accounting for full-node maximum-power consumption.

What Models Does BioIR Support?

BioIR is not limited to a single protein-folding architecture. The engine accelerates three major models: Boltz-2, OpenFold2, and OpenFold3. Developers can use BioIR in two ways. The end-to-end processor handles the entire workflow from raw input to final protein structure files in PDB or mmCIF format. Alternatively, developers can integrate selected optimized modules directly into custom PyTorch code, giving researchers flexibility to build specialized pipelines.

The software is available now as an open-source GitHub repository with precompiled components. Users need Python 3.12 or later, a compatible NVIDIA GPU with appropriate drivers, a staged model checkpoint, and pre-computed multiple sequence alignments (MSAs) for each protein chain. Notably, BioIR does not require compilation tools like nvcc, CUDA source code, CMake, or the full CUDA toolkit, making deployment straightforward for research teams.

Why Does This Matter for Structural Biology?

The shift from single-protein predictions to proteome-scale analysis represents a fundamental change in how researchers approach structural biology. Computational speed directly enables new science. Faster folding means researchers can screen more protein variants, explore protein complexes more thoroughly, and integrate structural predictions into drug discovery pipelines more efficiently. The energy efficiency gains also matter for large-scale research institutions and biotech companies running continuous folding campaigns.

The integration of protein-structure prediction with genomic analysis further amplifies the impact. Google's recent AI genome system evaluates 89 million potential single-nucleotide variants across the human genome and correlates these genetic changes with three-dimensional protein structures using AlphaFold to predict functional impact. This type of large-scale structural analysis would be impractical without the kind of throughput improvements BioIR delivers.

What Are the Practical Limitations?

NVIDIA is transparent about the scope of its benchmarks. The 2.90x speedup and energy estimates are specific to the tested dataset and hardware configuration. They exclude preprocessing steps like multiple sequence alignment generation, CPU resource allocation, data storage, data transfer, and retry logic. The company explicitly warns against generalizing these numbers to all BioIR-supported models or different datasets. Additionally, BioIR's Ray-based scaling distributes independent targets across GPUs but does not split a single protein-folding prediction across multiple GPUs. Context-parallel folding is planned but not yet available.

The processor supports ligand structure prediction, meaning it can predict how small molecules bind to proteins, but it does not currently predict ligand-affinity scores, which measure binding strength. Templates for structure prediction can be supplied manually, but BioIR does not run HHsearch or HMMsearch, the standard tools for finding evolutionary templates.

Steps to Deploy BioIR in Your Research Pipeline

• Prepare Your Data: Gather protein sequences and pre-compute multiple sequence alignments (MSAs) for each chain using standard tools; BioIR accepts both paired and unpaired MSAs for multi-chain inputs.
• Install Dependencies: Set up Python 3.12 or later, ensure you have a compatible NVIDIA GPU with current drivers, and download the BioIR wheel from the GitHub repository without needing to compile CUDA code.
• Choose Your Integration Path: Decide whether to use the end-to-end processor for complete workflow automation or integrate specific optimized modules into custom PyTorch code for specialized applications.
• Configure Ray Scaling: If processing large worklists, set up Ray to distribute independent targets across available GPUs on your node, ensuring CPU stages overlap with GPU computation.
• Monitor Output: Verify that folded structures are written to PDB or mmCIF files and validate results against your quality thresholds before scaling to production workloads.

The release of BioIR reflects a broader trend in computational biology: as AI models for protein structure and genomic analysis mature, the engineering challenge shifts from model accuracy to deployment efficiency. For research institutions and biotech companies running large-scale structural prediction campaigns, tools like BioIR can reduce both computational cost and time-to-results, enabling faster iteration in drug discovery and structural genomics projects.