Nvidia unveiled its BioNeMo Agent Toolkit, which provides domain-specific tools and skills for the agentic life sciences era.
Including more than a decade’s worth of Nvidia life sciences libraries, tools and open models, the toolkit enables AI agents, scientists and labs to work together by gathering evidence, reasoning across findings, running computational experiments and recommending the next best steps to accelerate discovery. Here, the agent is tasked with creating AI specifically for scientific advances.
It gives any agent or AI platform — from general-purpose assistants to specialized scientific agents, software platforms and in-house biopharma systems — the tools needed to synthesize and summarize scientific knowledge, call models, evaluate results, reason and execute next actions.
Nvidia made the announcement during the Bio event, which brings together 20,000 industry leaders in biotech from around the world.
Kimberly Powell, vice president of healthcare at Nvidia, said in a press briefing, “Science has always advanced when scientists get a better instrument, and the microscope changed the questions we could ask. The same is true for X-ray crystallography and gene sequencing, and now AlphaFold life sciences is the largest domain of science and the most consequential to humanity.”
She added, “It is how we discover the medicines that treat cancer, how we understand the genetics of disease, how we develop the vaccines that protect populations, and how we feed a growing world. AI is the new class of scientific instrument, and it has arrived in several ways, each one more powerful than the last.”
The toolkit includes Nvidia BioNeMo and is powered by Nvidia NIM microservices, Nvidia Parabricks, Nvidia NeMo and Nvidia Nemotron technologies, along with accelerated computing and skills — providing an open and trusted foundation for agentic life sciences.
More than 50 leading companies are already using it to advance scientific discovery, tapping into agent-callable skills for tasks including protein structure prediction, molecular docking, generative chemistry, genomic analysis, protein design and biomarker discovery.
“Frontier models are the brains. BioNeMo is the scientific toolbox. Together, they give AI
agents the skills of a PhD research assistant and the speed of a supercomputer,” said Jensen Huang, founder and CEO of Nvidia, in a statement. “For the first time, researchers can build AI agents that understand scientific knowledge, use scientific tools and execute scientific workflows. This is a new way to do science — one that can dramatically accelerate discovery across biology, chemistry, genomics and medicine.”
Open model and research organizations — including the Arc Institute, Open Molecular
Software Foundation and the University of Washington’s Institute for Protein Design — are
working with Nvidia to leverage BioNeMo to advance frontier models and make them more accessible through agent-ready workflows. The IPD collaboration has accelerated runtimes for state-of-the-art biodesign models like RosettaFold3, resulting in two times faster performance than the prior-generation mode and many additional applications to accelerate protein design efforts are ongoing, giving researchers tools at a scale and cost not before possible.
“Every tool we’ve built for protein design is only as powerful as the scientists who can
efficiently access it,” said David Baker, professor of biochemistry at the University of
Washington School of Medicine and director of the Institute for Protein Design, in a statement. “The next leap in science won’t come from a single discovery; it will come from the speed of iterative designs and agents that can repeatedly reason through the complexity of biology at a
speed humans never could.”
Agent-Ready Tools and Skills for Life Sciences
Life sciences is one of the world’s most important scientific frontiers, with global scientific
R&D reaching $3.8 trillion and annual pharmaceutical budgets approaching $300 billion.
Powell said that generative AI models that could create protein sequences, molecules, hypotheses, and then reasoning AI models that don’t just generate, they think and break down complex things into a step-by-step reasoning.
“And now the agent harnesses an OpenClaw, gives a model memory, tools, skills, and a loop, and [the model] stops being something that says things, and it starts being something that does things, and developers recognize this pattern immediately…. The $300 billion life sciences R&D industry is at an inflection point. The question is no longer whether AI can help science. The question is, ‘Does AI have the right instruments to run science?'”
Agentic workflows can help the industry iterate faster while reducing costs and maximizing
the probability of success. With the toolkit allowing developers to transform general-purpose agents into life sciences agents in minutes, researchers can run experiments faster, continuously learn from results and close the loop between hypothesis and discovery, with some companies extending this iteration into physical labs.
Nvidia is optimizing the entire BioNeMo platform by turning libraries, models and
frameworks into agent-callable tools. This includes harnessing Nvidia Agent Toolkit
technologies such as Nvidia Nemotron for the reasoning foundation, the Nvidia NeMo RL
library for reinforcement learning and Nvidia NemoClaw blueprints for secure, private
agents that can reason across tasks, call tools and interact with data continuously.
Nvidia NIM microservices help agents call models and perform tasks. The Nvidia
OpenShell runtime provides a controlled executable environment.
Powell said, “AI for science is a full stack computing problem. At the bottom, we still need energy chips, infrastructure, and networking. This is the foundation everything runs on above the model layer, and for the first time industries can codify decades of scientific expertise directly into models that they own and control. That’s why Nvidia builds open models at the frontier of AI, open language models, biology models, chemistry models, and world models, giving model builders a proven foundation of open weights and open architecture and data they can extend with their own proprietary data and scientific IP, retaining full ownership of what they build.”
The toolkit’s components enable agents to complete workflows such as:
● Virtual Screening: Agents can help researchers identify small-molecule drug candidates by generating and screening compounds, docking them to a target, predicting binding strength and filtering for drug-like properties. Then, the agent can output which candidates should be prioritized, compressing screening timelines from days to minutes.
● Genomic Analysis and Target Discovery: Agents can help researchers transform raw sequencing data into prioritized genetic insights and biological targets. Nvidia Parabricks accelerates alignment and variant calling, while genomic foundation models score variant effects and the agent ranks the most disease-relevant candidates for further study.
● Protein Binder Design: Agents can help researchers design and validate candidates
computationally before work begins, compressing traditionally labor-intensive design work.
● Deep BioMedical Research: Agents connect real-world data to reasoning models to improve the efficiency and accuracy of various scientific and clinical development processes, including literature review, protocol generation, clinical trial screening and pharmacovigilance with the BioMedical AI-Q Research Agent.
● Medical Imaging Analysis: Agents can help researchers process, segment, synthesize and reason over medical imaging data to support biomarker discovery, accelerating evidence generation across research workflows.
Life Sciences Ecosystem Builds With Nvidia BioNeMo
Companies across the technology and life sciences ecosystem are using the toolkit to
advance agentic workflows.
Frontier labs and scientific agent builders including Edison Scientific, Lila Sciences, OpenAI
and Owkin are using BioNeMo to help agents move from answering questions to completing scientific work. Nvidia accelerated models and analysis libraries help shorten the time from hypothesis to insight.
Scientific data and workflow platforms from Benchling, Certara, Databricks, Snowflake and
Seqera are using BioNeMo Agent Toolkit to connect data systems with AI-powered science.
BioNeMo skills can help agents query biological and chemical datasets, prepare model-ready inputs, launch reproducible workflows, analyze outputs and return insights directly within the platforms scientists and data teams already use daily.
Diagnostics and pharmaceutical companies including Lilly and Natera are using BioNeMo Agent Toolkit to scale my repeatable agentic workflows across discovery, translational research and clinical insight.
AI-native biology companies including Boltz, Basecamp Research, Chai Discovery, PerturbAI,
Dyno and Proxima have collaborated with Nvidia to develop tools to accelerate model-powered therapeutic design workflows.
Computer-aided drug discovery software providers including Dassault Systèmes, Cadence
(OpenEye) and Schrödinger are integrating the toolkit’s capabilities into scientific applications used across discovery teams. Then, agents can help orchestrate molecular generation, docking and prediction, turning computer-aided design platforms into systems where researchers can ask questions, launch analyses and identify next-best actions faster.
Lab instruments and automation companies including Automata, HighRes,Tecan and Thermo Fisher and autonomous data generation platform, Medra are connecting systems with computational discovery powered by BioNeMo skills.
AI clouds and AI infrastructure companies including Baseten, Modal and Nebius are using
the toolkit to help developers build life sciences workflows and reliable hosted services. By
supporting BioNeMo skills and tools through scalable application programming interfaces,
managed compute and production inference environments, these companies can help move agentic biology workflows from prototypes into accessible services for researchers and enterprises.
Availability
BioNeMo Agent Toolkit and skills are available now through the NVIDIA developer resources
page and GitHub.