NVIDIA BioNeMo Inference Runtime is live in public beta as of today.
#OpenFold3 developed by @MoAlQuraishi and OpenFold Consortium now runs >2× faster on H100 (and other) GPUs thanks to specialized optimizations that general-purpose compilers miss.
👋 Say hello to faster structure prediction
NVIDIA BioNeMo Inference Runtime is in public beta. This open, PyTorch-native library speeds up biomolecular inference on NVIDIA GPUs.
Optimize models including Boltz-2, OpenFold2, and Protenix v2 with specialized kernels and, where supported, CUDA Graphs - then scale with Ray-powered GPU replicas.
🙌 Huge thanks to our collaborators at @apheris_AI, @AurekaBio, @Biomap_IMI, @boltz_bio, @dataiku, @UWproteindesign, @Latent_Labs, MoleculeMind, @nebiusai, @openmsf, @open_fold, @proximabio, @SandboxAQ, @Terray_Tx, and @Xaira_Thera for partnering with us on this work.
📃 nvda.ws/46PDmlL
openfold retweeted
For my first post, I’m sharing a letter @NVIDIA signed on why open models matter.
AI will transform every industry, power every company, and be built by every country.
Open models strengthen safety and cybersecurity, accelerate innovation and diffusion, and enable sovereignty.
The world needs both frontier closed models and frontier open models.
images.nvidia.com/pdf/Open-W…
openfold retweeted
Biomolecular structure prediction is a full-stack challenge, not just a model benchmark.
See how NVIDIA accelerates structure prediction with GPU MSA search up to 177× faster, OpenFold3 inference up to 4× faster on NVIDIA Blackwell GPUs, and Fold-CP scaling to 32,000-token complexes across 64 NVIDIA B300 GPUs, composed into agentic workflows with BioNeMo Agent Toolkit. 🧬
Read the full blog nvda.ws/4yfAn2o
Full article on Business Wire: businesswire.com/news/home/2…
bioRxiv paper: biorxiv.org/content/10.64898…
Download the data: docs.google.com/forms/d/e/1F…
Congratulations to Gabriel Rocklin, Kotaro Tsuboyama, Yehlin Cho, Sergey Ovchinnikov, and everyone involved in creating this important resource.
This is core to the OpenFold mission: open foundation models, open data, open tools, and a collaborative scientific community working to democratize AI for biology and drug discovery.
Structure prediction has transformed the field, but structure alone is not enough. The next generation of biomolecular AI must learn biophysical properties: stability, folding, aggregation risk, designability, molecular recognition, and more.
In biology, public datasets often overrepresent successful examples: stable structures, functional proteins, binders, designs that worked. But for machine learning, the failures are not noise. They are essential signal. Models also need to see what does not fold.
It also matters because this dataset includes something the field badly needs more of: negative data.
Protein stability is one of the foundational properties in biology. It determines whether proteins fold, remain functional, avoid aggregation, and can be engineered into useful tools, therapeutics, enzymes, and biological systems.
openfold retweeted
We @open_fold are excited for the first major data release from @grocklin partly funded by our consortium. First but by no means the last! Large scale, high quality, diverse stability datasets like this are crucial for more useful protein #ai for #biotech businesswire.com/news/home/2…
In keeping with our commitment to open science, all resulting code will be released under permissive licenses so researchers and companies worldwide can build on it. Read more in our linked press announcement! businesswire.com/news/home/2…
