{"id":2135,"date":"2026-07-10T13:00:00","date_gmt":"2026-07-10T13:00:00","guid":{"rendered":"https:\/\/futurenews24.com\/index.php\/2026\/07\/10\/accelerating-end-to-end-co-folding-performance-with-nvidia-bionemo-agent-toolkit\/"},"modified":"2026-07-10T13:59:11","modified_gmt":"2026-07-10T13:59:11","slug":"accelerating-end-to-end-co-folding-performance-with-nvidia-bionemo-agent-toolkit","status":"publish","type":"post","link":"https:\/\/futurenews24.com\/index.php\/2026\/07\/10\/accelerating-end-to-end-co-folding-performance-with-nvidia-bionemo-agent-toolkit\/","title":{"rendered":"Accelerating Finish-to-Finish Co-Folding Efficiency with NVIDIA BioNeMo Agent Toolkit"},"content":{"rendered":"
\n<\/p>\n
\n
Biomolecular construction prediction and co-folding with fashions like OpenFold3 at the moment are mainstream, large-scale workloads powering drug discovery and protein design. More and more, they\u2019re pushed end-to-end by AI brokers. For an agent to run that pipeline effectively, each step must be quick and scalable: A number of Sequence Alignment (MSA) era, co-folding inference, serving, and multi-GPU scale-out. A bottleneck anyplace limits total throughput.\u00a0<\/p>\n
Velocity and memory-efficiency are important for key drug discovery workflows similar to digital screening and prediction of enormous molecular assemblies. In digital screening, hundreds of thousands to billions of compounds are screened in opposition to one or a couple of protein targets. Whereas co-folding fashions usually give the most effective predicted constructions, they are often costly to run, making them impractical for digital screening functions. That’s the place NVIDIA acceleration turns into key, making potential the deployment of OpenFold3 and associated strategies on the scale of enormous compound libraries.<\/p>\n
Velocity can be necessary for predicting giant molecular assemblies involving a number of proteins and hundreds of amino acid residues, as co-folding mannequin runtime scales cubically with the variety of residues. A fair greater problem, nonetheless, is reminiscence use, as single GPU reminiscence may be restricted, putting a tough ceiling on the scale of complexes that may be predicted in a single shot. Strategies to scale back reminiscence necessities, and that distribute prediction duties throughout a number of GPUs, would allow qualitatively new functions which might be merely unfeasible at present.<\/p>\n
NVIDIA has constructed instruments to speed up and enhance the effectivity of every step of the construction prediction and co-folding workflow. NVIDIA BioNeMo Agent Toolkit offers brokers seamless entry to the instruments they should speed up biology and chemistry workflows. On this put up, we break down the accelerations for every stage on NVIDIA B300 and H100 GPUs, then present how these phases may be executed by an agent (see Determine 1, beneath).<\/p>\n
\nDetermine 1. The NVIDIA-accelerated structure-prediction workflow: GPU MSA, cuEquivariance, optimized inference, and Fold-CP, all orchestrated by the BioNeMo Agent Toolkit<\/figcaption><\/figure>\n<\/div>\n
Take away the MSA bottleneck with GPU MSA<\/h2>\n
For co-folding fashions, constructing the MSA has historically been a CPU-bound step that may dominate wall-clock time. MMseqs2-GPU strikes homology search onto NVIDIA GPUs, lowering this bottleneck whereas scaling with sequence size on each NVIDIA Hopper and NVIDIA Blackwell architectures. <\/p>\n
The most recent GPU accelerated model provides Hopper and Blackwell particular optimizations, together with environment friendly assist for larger-than-GPU-memory database search on NVIDIA Grace programs and extra speedups from improved Blackwell DPX directions obtainable from CUDA 13.2. These GPU contributions have been upstreamed again into the principle MMseqs2 repository so the entire group can profit from the accelerations.<\/p>\n
The MSA Search NIM makes use of MMseqs2-GPU, whose Nature Strategies paper reviews as much as 177\u00d7 quicker alignment than CPU JackHMMER on a single L40S. In our benchmarking, the stage scales easily previous 10k tokens on H100 and B300 GPUs (see Determine 2, beneath). The MSA Search NIM may be referred to as instantly, self-hosted or wrapped as a software in an agentic workflow.\u00a0<\/p>\nNotice: We present learn how to name the NIM API endpoint hosted on construct.nvidia.com, however for large-scale workloads, the self-hosted NIM is the popular beneficial endpoint. construct.nvidia.com endpoints aren’t constructed to deal with giant quantity requests and can time-out.<\/figure>\n
\n# Add the MSA Search NIM talent to your agent (browse all: add –list) \nnpx abilities add NVIDIA-BioNeMo\/bionemo-agent-toolkit –skill msa-search-nim –agent claude-code \n# Use hosted API on construct.nvidia.com (nothing to obtain) \n# It’s also possible to use the talent to obtain the NIM container and supply self-hosted API endpoint. We do not cowl that on this tutorial. \nexport NVIDIA_API_KEY=<\/p>\n
# Simply immediate the agent: \n# It’s a must to obtain a pattern goal.fasta file. You may immediate the agent to obtain it for you or level to an already present file. \n“Construct an MSA for the sequence in goal.fasta with the MSA Search NIM.”\n<\/p><\/div>\n
\nDetermine 2. MSA-stage latency scales close to linearly to 10k+ tokens on H100 and B300 (429 s vs 463 s at ~10k tokens)<\/figcaption><\/figure>\n<\/div>\n
Fold at SOTA velocity with cuEquivariance and the OpenFold3 NIM<\/h2>\n
cuEquivariance is a CUDA-X library of geometric studying primitives for atomistic modeling and it gives accelerated variations of the Triangle Consideration, Triangle Multiplication and Consideration Pair Bias kernels that dominate co-folding. On B300 it cuts latency as much as ~3\u00d7 (see Desk 1, beneath).\u00a0<\/p>\nSequence lengthPyTorch (OSS)cuEquivarianceSpeedup1,024 (H100)79.4 s41.6 s1.9\u00d71,024 (B300)49.1 s27.3 s1.8\u00d71,536 (B300)149.0 s56.2 s2.7\u00d72,048 (B300)300.1 s97.6 s3.1\u00d7Desk 1. Comparability of OpenFold3 forward-pass latency with and with out cuEquivariance<\/figcaption><\/figure>\n
cuEquivariance kernels are built-in instantly into the OSS fashions like OpenFold3 (offered as an non-compulsory dependency), OpenFold2, RosettaFold3, Protenix and Boltz. <\/p>\n
As a result of the accelerations are upstreamed into these OSS fashions, a researcher will get the speedups mechanically just by working the mannequin they already use on an NVIDIA GPU. CuEquivariance kernels additionally extends most sequence size to ~5.9k tokens whereas PyTorch runs out of reminiscence past ~1.5k\u20132.5k tokens.<\/p>\n
\nDetermine 3. cuEquivariance lowers latency and extends most sequence size over PyTorch on H100 and B300<\/figcaption><\/figure>\n<\/div>\n
On high of cuEquivariance, the OpenFold3 NIM applies additional inference optimizations that compound the achieve (see Determine 4, beneath), attaining sequence lengths of as much as ~6,400 on a single B300. These extra accelerations are delivered by the NIM; for SOTA out of the field, builders can name the NIM endpoint instantly or compose it into an agentic workflow.<\/p>\n
\nDetermine 4. The totally optimized OpenFold3 NIM on H100 and B300, attaining over 3x and 4x quicker, respectively, folding as much as 6,400 tokens on a single B300\u00a0<\/figcaption><\/figure>\n<\/div>\n
\n# Add the OpenFold3 NIM talent to your agent \nnpx abilities add NVIDIA-BioNeMo\/bionemo-agent-toolkit –skill openfold3-nim –agent claude-code<\/p>\n
# Hosted API on construct.nvidia.com \n# It’s also possible to use the talent to obtain the NIM container and supply self-hosted API endpoint. We do not cowl that on this tutorial. \nexport NVIDIA_API_KEY=<\/p>\n
# Immediate the agent: \n# It’s a must to obtain a pattern goal.fasta file. You may immediate the agent to obtain it for you or level to an already present file. \n“Fold goal.fasta with OpenFold3 utilizing the MSA from the earlier step; return the ranked constructions with confidence scores.”\n<\/p><\/div>\n
Scale past one GPU with Fold-CP<\/h2>\n
Single-GPU reminiscence has traditionally capped co-folding fashions at a couple of thousand residues. On NVIDIA B300 (Blackwell Extremely), the bigger HBM and Blackwell-generation effectivity mixed with the cuEquivariance and superior inference optimizations above push that ceiling considerably larger with no mannequin adjustments.\u00a0<\/p>\n
In lots of instances, being single-device certain is probably not ample. Fold-CP introduces a brand new parallelization method such that per-device reminiscence requirement scales as O(N\u00b2\/P) the place N is token depend and P is the variety of GPUs, reaching 32,000 tokens on 64 B300 with the Boltz-2 mannequin\u2014a couple of 12\u00d7 leap over the single-GPU restrict.\u00a0<\/p>\n
To strive Fold-CP, merely level your agent to the Boltz-CP codebase and ask it to run multi-GPU inference.\u00a0<\/p>\n
\n# Context-parallel inference throughout 4 GPUs with Fold-CP (boltz-cp) \ngit clone https:\/\/github.com\/NVIDIA-Digital-Bio\/boltz-cp && cd boltz-cp \n# Set up dependencies after which run the command beneath \ntorchrun –nnodes 1 –nproc_per_node 4 \n src\/boltz\/distributed\/important.py predict \/path\/to\/preprocessed_data \n –out_dir .\/predictions \n –size_dp 1 –size_cp 4 \n –recycling_steps 3 –sampling_steps 200 –diffusion_samples 5\n<\/div>\n
Accelerating the end-to-end co-folding pipeline<\/h2>\n
Construction prediction efficiency is now an end-to-end programs downside. For OpenFold3, the sensible workflow spans MSA era, co-folding inference, deployment, and the reminiscence limits that decide how giant a organic meeting may be modeled. <\/p>\n
NVIDIA accelerates every layer of that workflow: MSA Search NIM speeds homology search by 177x, cuEquivariance and OpenFold3 NIM decrease inference latency by as much as 4x on Blackwell GPUs, and Fold-CP reveals how context parallelism can lengthen co-folding past a single GPU to 32,000-token complexes on 64 NVIDIA B300 GPUs.\u00a0<\/p>\n
Collectively, these instruments make construction prediction quicker, extra scalable, and simpler to compose into agentic discovery workflows, serving to researchers transfer from mannequin predictions to bigger, extra helpful organic programs.<\/p>\n
What these accelerations unlock<\/h2>\n
These enhancements open up lessons of structural biology issues that have been beforehand out of attain. In digital screening, quicker co-folding inference implies that structure-based strategies, which have traditionally been reserved for the ultimate phases of a drug discovery marketing campaign, can now be utilized at far earlier phases and at a lot bigger library scales, enhancing the standard and variety of hits that advance by the pipeline. <\/p>\n
For giant biomolecular assemblies, the mix of prolonged single-GPU capability on B300 and the context-parallel Fold-CP framework shifts what’s modelable: complexes on the scale of the ribosome, the spliceosome, or giant signaling assemblies have been structurally intractable for co-folding fashions, and these accelerations start to alter that. To make this concrete, folding a posh of ~10,000 residues, roughly the dimensions of the bacterial ribosome, would have been prohibitively costly or just out of attain on a single GPU; on B300 with Fold-CP, such predictions change into tractable throughout a multi-GPU node. <\/p>\n
These are qualitative shifts within the questions that structural biology can ask computationally, not merely enhancements in throughput. Making these instruments accessible by open-source integrations and agentic APIs will speed up the tempo at which computational predictions translate into organic perception and, finally, into new medicines.<\/p>\n
Getting began\u00a0<\/h2>\n
Strive the accelerated OpenFold3 workflow with NVIDIA BioNeMo Agent Toolkit, beginning with these instruments:<\/p>\n
Acknowledgments\u00a0<\/h4>\n
We\u2019d prefer to thank our broader NVIDIA staff for growing the benchmarks and instruments:\u00a0Franco Pellegrini, Lalit Vaidya, Duc Tran, Tien Pham, Maximilian Stadler, Alejandro Chacon, Quan Vu, Simon Chu, Brian Roland, Dejun Lin, Joseph Chang, Hoa La, Jonathan Mitchell, Vishanth Iyer, Timur Rvachov, Christian Dallago, Christian Hundt. We\u2019d additionally prefer to thank the broader OpenFold and OMSF groups for our collaboration and their contributions.<\/p>\n<\/div>\n
![\"A](\"https:\/\/developer-blogs.nvidia.com\/wp-content\/uploads\/2026\/07\/image3-5.webp\")
![\"A](\"https:\/\/developer-blogs.nvidia.com\/wp-content\/uploads\/2026\/07\/image6-2.webp\")
![\"A](\"https:\/\/developer-blogs.nvidia.com\/wp-content\/uploads\/2026\/07\/image2-4.webp\")
![\"Side-by-side](\"https:\/\/developer-blogs.nvidia.com\/wp-content\/uploads\/2026\/07\/image4-5.webp\")