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

In protein co-folding pipelines, multiple sequence alignment (MSA) generation has traditionally been a CPU bottleneck, often dominating total runtime. NVIDIA has integrated a GPU-accelerated version of MMseqs2 (MMseqs2-GPU) into the BioNeMo Agent Toolkit, achieving up to 177x speedup over CPU JackHMMER on Hopper and Blackwell architectures (based on Nature Methods paper data on a single L40S). This acceleration has been upstreamed to the main MMseqs2 repository, benefiting the community directly.

MMseqs2-GPU scales smoothly to 10k+ tokens on H100 and B300; for example, at ~10k tokens, H100 latency is 429 seconds and B300 latency is 463 seconds, with near-linear performance. The tool is available as an MSA Search NIM, supporting direct calls or use as an agent tool. For large-scale workloads, NVIDIA recommends using self-hosted NIM endpoints rather than the hosted API on build.nvidia.com, which is not suitable for high-throughput requests.

cuEquivariance is NVIDIA's CUDA-X geometric learning primitive library, accelerating core kernels in co-folding models such as Triangle Attention, Triangle Multiplication, and Attention Pair Bias. On B300, cuEquivariance reduces OpenFold3 forward pass latency by up to 3.1x (from 300.1s to 97.6s at 2048 tokens) and extends the maximum sequence length from PyTorch's ~1.5k–2.5k tokens to ~5.9k tokens. On H100, the speedup is 1.9x at 1024 tokens (79.4s vs 41.6s).

OpenFold3 NIM builds on cuEquivariance with additional inference optimizations, achieving up to 6,400 tokens on a single B300 and overall speedup of over 4x compared to the original PyTorch implementation (over 3x on H100). These accelerations have been upstreamed to open-source models including OpenFold3, OpenFold2, RosettaFold3, Protenix, and Boltz; users simply run existing models on NVIDIA GPUs to automatically benefit.

Co-folding models have cubic computational complexity with respect to residue count, and single-GPU memory limits impose a hard cap on complex size. NVIDIA's Fold-CP technology uses context parallel inference to reduce per-GPU memory to O(N/P) (N = sequence length, P = number of GPUs), enabling 32,000-token mega-complex modeling on 64 B300 GPUs, covering ribosome-scale molecular machines.

This breakthrough makes previously infeasible structural biology problems (e.g., end-to-end prediction of large multi-protein complexes) possible, providing key capabilities for virtual screening and large complex prediction in drug discovery. Fold-CP is also integrated into the BioNeMo Agent Toolkit, supporting multi-GPU scaling in agent workflows.

All accelerated components—MMseqs2-GPU, cuEquivariance, OpenFold3 NIM, and Fold-CP—are orchestrated through the NVIDIA BioNeMo Agent Toolkit. The toolkit allows developers to add NIM skills to agents (e.g., Claude Code) via simple command-line instructions (e.g., npx skills add), and supports both hosted API and self-hosted NIM deployment.

NVIDIA emphasizes that end-to-end acceleration is critical for large-scale virtual screening: traditionally, co-folding models are too computationally expensive to apply to billions of compounds, but full GPU pipeline acceleration makes it feasible. Additionally, cuEquivariance and MMseqs2-GPU accelerations have been upstreamed to open-source repositories, ensuring continued community benefit.