# Accelerating End-to-End Co-Folding Performance with NVIDIA BioNeMo Agent Toolkit

[NVIDIA Developer Blog](https://yomu.fyi/company/nvidia-developer-blog) · Elizabeth Goodman · Jul 10, 2026

**Type:** Problem & solution

## Summary

Biomolecular structure prediction and co-folding pipelines encounter latency and memory ceilings during sequence alignment and multi-residue inference. NVIDIA addresses these constraints through the BioNeMo Agent Toolkit, which connects AI agents to accelerated biological computing components. The pipeline offloads multiple sequence alignment to MMseqs2-GPU, while cuEquivariance and the OpenFold3 NIM accelerate core geometric attention kernels to cut inference latency up to fourfold. To handle large molecular assemblies beyond single-GPU memory limits, the Fold-CP framework distributes context across multiple GPUs with O(N²/P) memory scaling. Together, these tools enable the prediction of complexes up to 32,000 tokens on 64 NVIDIA B300 GPUs, making large-scale virtual screening and ribosome-scale modeling computationally feasible.

## Context

Biomolecular structure prediction and co-folding workloads like OpenFold3 face severe throughput and memory bottlenecks in agentic drug discovery and protein design pipelines. Multiple Sequence Alignment generation has traditionally been CPU-bound, while co-folding runtime scales cubically with residue count. Single-GPU memory limits also cap the size of complexes that can be predicted in a single pass.

## Approach / What changed

NVIDIA created an end-to-end acceleration workflow within the BioNeMo Agent Toolkit. Homology search is accelerated via MMseqs2-GPU in the MSA Search NIM. Inference relies on cuEquivariance kernels and the OpenFold3 NIM for optimized execution. For massive assemblies, the Fold-CP framework distributes memory requirements across multiple GPUs using context parallelism.

## Takeaways

- MMseqs2-GPU moves Multiple Sequence Alignment to GPUs, achieving up to 177x faster alignment than CPU JackHMMER on an L40S and scaling past 10k tokens on H100 and B300 GPUs.
- The cuEquivariance library accelerates Triangle Attention, Triangle Multiplication, and Attention Pair Bias kernels, cutting OpenFold3 latency by up to 3.1x on B300 compared to PyTorch.
- Fold-CP distributes per-device memory at O(N²/P), enabling the Boltz-2 model to scale to 32,000 tokens across 64 NVIDIA B300 GPUs.

**Tags:** [Machine Learning](https://yomu.fyi/topic/machine-learning), [Open Source](https://yomu.fyi/topic/open-source), [Performance](https://yomu.fyi/topic/performance), [Scalability](https://yomu.fyi/topic/scalability)

[Read original post](https://developer.nvidia.com/blog/accelerating-end-to-end-co-folding-performance-with-nvidia-bionemo-agent-toolkit)