When biomedical research outgrows local compute

How the University of Arizona’s ANBM Center built a hybrid GPU pipeline for molecular simulation and gas-based therapeutics

At the University of Arizona’s Center for Applied NanoBioscience and Medicine, researchers are studying whether gases such as xenon and hydrogen could be developed into new forms of therapy.

The center has created formulations known as “gas marbles.” These use nanoparticles and liquid films to stabilize small quantities of gas so they can potentially be delivered to targeted tissue. The team is investigating several delivery routes, including topical, intranasal, oral, and systemic administration.

One part of the program is assessing noble gases as possible radiomitigators: treatments intended to reduce harm following exposure to ionizing radiation. The research could eventually have applications in radiation emergencies, radiotherapy, severe burns, and exposure to cosmic radiation during deep-space flight.

Before any candidate advances to laboratory validation, the researchers need to understand how individual gases might interact with proteins and which interaction sites deserve closer study.

That turns a biomedical question into a substantial computing problem.

The compute problem starts before the laboratory

The interactions involved are difficult to observe directly. Researchers therefore use computational models to predict protein structures, locate possible interaction sites, and simulate how gas atoms and molecules behave around them over time.

The ANBM Center is building a pipeline that combines AI-based structure prediction, molecular docking, AMBER molecular dynamics, and machine-learning analysis.

Each method answers a different question. Structure-prediction tools provide a model of the protein. Molecular docking helps the team identify promising sites where a gas might interact. Molecular dynamics then models how atoms move and influence one another over time, including weak forces such as van der Waals interactions.

The center is also exploring whether effects such as atomic spin could prove relevant in some interactions. That remains an open research question rather than an established mechanism.

Greater computing capacity does not make a scientific model correct by itself. It can, however, let researchers examine more candidates, run longer or repeated simulations, and test whether results remain consistent under different conditions.

“High-quality capacity GPU allows us to improve the accuracy of our models and accelerate cycle time from days to minutes,” said Frederic Zenhausern, director of the ANBM Center.

“Both time and accuracy are crucial to our data.”

‍

‍

Local resources reached their limit

The center began the program with limited PC-based computing available through the College of Medicine–Phoenix. It gradually added a Mac Mini M3 Max, a GPU workstation, and then a small multi-node GPU cluster.

That local infrastructure remains part of the research setup. It was never intended to carry every stage of the growing pipeline.

Capacity and availability became the main constraints as the team began working with larger simulations. Zenhausern said access to the university’s central high-performance computing resources in Tucson had been unreliable for the center’s needs. The researchers also needed direct access to the infrastructure so routine administrative requests or hardware maintenance would not delay the work.

The answer was not to abandon local computing. The center added external GPU capacity for the stages where the local environment was no longer sufficient.

A hybrid pipeline rather than a full migration

The current workflow distributes work according to what each environment can handle efficiently.

Researchers develop their methods and perform initial molecular docking on local workstations and the center’s own cluster. This first screening stage identifies the protein–gas interaction sites that appear most promising.

The selected candidates then move to Hivenet for larger AMBER molecular-dynamics simulations and, where appropriate, machine-learning analysis.

The resulting data return to the research team for interpretation. Researchers inspect the outputs, check the simulation parameters, compare findings, and decide which candidates should progress to further computational work or future laboratory studies.

This division keeps early experimentation close to the researchers while giving them additional capacity when a workload becomes too large for the local setup.

It also avoids treating cloud infrastructure as the default destination for every task. The center uses it as one layer of a broader research environment.

US-based operations were a selection requirement

The location and operation of the infrastructure mattered alongside capacity, cost, and access.

“US-based infrastructure and operations were critical to our selection, especially due to governmental compliance required by NIH and other federal agencies,” Zenhausern said.

The center treated this as a hard requirement for the relevant workload. The practical question was where the computation would run and where the service supporting it would operate.

That is more specific than a general claim about cloud sovereignty. For ANBM, control meant selecting US-based capacity that could fit the institutional and federal obligations attached to the program.

The center already maintains relationships with several large technology providers. Hivenet did not replace those partnerships. It was added to the center’s infrastructure mix for a particular research need, based on the combination of US operations, available capacity, cost, and direct technical support.

Getting researchers onto the infrastructure

From the signed agreement to a researcher running a workload took less than two months. Zenhausern said much of the elapsed time came from processes on the university side, while Hivenet handled its part of the setup quickly.

For the center, onboarding involved more than granting access to a GPU. The environment had to become usable by graduate students and other researchers without asking each person to learn how to maintain the underlying infrastructure.

The team now works through standardized project environments, shared scripts, and documented procedures. Graduate students can run routine computational tasks, organize outputs, and work with senior researchers to validate parameters and interpret results.

That structure also supports reproducibility. Researchers can see which procedures and settings produced a result instead of relying on an undocumented setup maintained by one specialist.

“The relationship has been closer to a technical partnership than a conventional vendor arrangement,” Zenhausern said.

“Hivenet’s team has been responsive to our research requirements, listened to how our workflows operate, and helped us think through infrastructure needs as the program develops.”

Early gains, with limits on what can be claimed

The additional capacity has allowed the center to explore more concepts and screen a larger set of potential candidates before deciding which ones merit experimental study.

That could help the team design narrower in-vitro experiments, reduce unnecessary use of laboratory reagents, and eventually limit the number of candidates that need to advance to animal studies. These remain expected benefits of the pipeline rather than measured outcomes from a completed research program.

Zenhausern described some computational cycles as moving from days to minutes. He also said that it is too early to provide a precise assessment of simulation time, cost per run, throughput, or the number of concurrent jobs compared with the previous setup.

The center has not yet identified a publication or scientific milestone that it is ready to associate publicly with the partnership.

The clearest result at this stage is therefore operational. ANBM can run a broader computational program than its original local setup allowed, while developing the scientific methods and measurement practices needed to evaluate that program properly.

The access problem goes beyond hardware

Expanding computational research creates its own institutional pressure. The tools require funding, specialist knowledge, and people who can operate them.

“It is great to have more tools to do our research, but it is also somewhat scary, as these tools require financial resources, knowledge and talents which may not be readily available,” Zenhausern said.

A GPU may be available in principle and remain inaccessible in practice. A research center still needs a budget, an approved provider, usable environments, documented procedures, and support when a workload changes or fails.

ANBM’s experience suggests that access should be measured by what researchers can run, rather than the hardware an institution can technically point to.

The center is now seeking sustainable resources to continue developing its computational pipeline and make external capacity a dependable part of its research infrastructure.

What the University of Arizona case shows

The ANBM Center did not move its entire research environment into the cloud. It built a hybrid pipeline around the way its scientists already worked.

Early screening remains local. Larger simulations use external GPU capacity. Interpretation and experimental planning return to the research team.

That arrangement gave the center direct access to additional compute without requiring graduate students to become hardware operators or the center to build a dedicated infrastructure team.

It also made infrastructure constraints part of the design from the beginning. Capacity, US-based operations, institutional processes, reproducibility, and technical support all influenced what the center could use.

The scientific outcome of the gas-therapeutics research remains open. The infrastructure outcome is already clearer: the team can ask larger computational questions and move promising candidates through a more capable research pipeline.

For research centers building similar workflows, that may be the most useful measure of access.

‍