NVIDIA Unveils BioNeMo Inference Runtime (BioIR) to Accelerate Biomolecular Structure Prediction

NVIDIA has introduced the BioNeMo Inference Runtime (BioIR), a new Python library engineered to dramatically accelerate the performance of supported biomolecular structure prediction models when executed on NVIDIA GPUs. This development is particularly timely as the field of biomolecular structure prediction transitions from analyzing individual protein targets to undertaking proteome-scale worklists. The primary computational bottleneck has shifted from the inherent capability of a model to fold a protein to the sheer speed at which an entire queue of independent targets can be processed through the sequential stages of parsing, featurization, GPU inference, and output writing. BioIR is designed to optimize this entire workflow while seamlessly integrating with the standard PyTorch environment, ensuring that models remain ordinary `torch.nn.Module` objects. This approach eliminates the need for complex engine building, export steps, or the creation of separate artifacts between a model checkpoint and its forward pass, thereby simplifying the development and deployment pipeline.

BioIR has already demonstrated its efficacy at a production scale, notably playing a crucial role in the recent expansion of the AlphaFold Database. This expansion involved the generation of protein-complex structures across an impressive 4,777 proteomes, leading to the analysis of approximately 31 million candidate complexes. From this vast dataset, 1.81 million predictions were identified and released as high-confidence structures. The BioIR library is now readily available to the research community as an open GitHub repository, distributed as a Python wheel containing precompiled CUBINs for ease of installation and use. To successfully deploy BioIR, users will require Python version 3.12 or later, a compatible NVIDIA GPU and its associated driver, a staged model checkpoint, and per-chain A3M MSAs (Multiple Sequence Alignments). A significant advantage of BioIR is its reduced dependency, as it does not require the installation of `nvcc`, CUDA source code, CMake, or the full CUDA toolkit, thereby streamlining the setup process.

The core optimization strategy of BioIR targets specific operations that are not fully addressed by more general-purpose inference stacks. This includes specialized components crucial for protein structure prediction, such as Pairformer and Evoformer stacks, triangle operations, pairwise attention mechanisms, diffusion transformers, and atom-level modules. The library offers two primary modes of utilization. The first is an end-to-end processor, which manages the complete pipeline from an `InputRequest` through parsing, tokenization, feature generation, GPU inference, and finally, writing the output in standard formats like PDB or mmCIF. The second mode provides direct PyTorch integration, enabling developers to construct supported models from scratch or to reuse specific optimized modules within their own custom codebases. A practical tutorial is available that demonstrates the processor path using the Boltz-2 model, with the model source specified as `boltz_2`. Each protein chain processed by BioIR necessitates an A3M MSA. The library supports both paired and unpaired MSAs for inputs that involve multiple non-identical protein chains, and also offers the flexibility to incorporate user-supplied templates.