Academic medical centers are at the frontier of a shift that is redefining life science research. AI models like AlphaFold3 are opening new doors in drug discovery, precision medicine, and our fundamental understanding of biological systems, but deploying them at institutional scale requires container orchestration, GPU resource management, secure data handling, and downstream visualization pipelines that most research teams simply don't have.
At UT Southwestern Medical Center, one of the nation's leading academic medical centers, the research computing team built Astrocyte, an innovative research computing platform to solve that problem.
But a bigger challenge remained. Each new model release or new version upgrade or emergent AI capability would trigger the same manual cycle: containerization, testing, documentation, user onboarding. The UTSW infrastructure team that successfully deployed one tool today would face the same sprint for the next one tomorrow.
Now, in collaboration with IBM and NVIDIA, they're tackling this bigger challenge.
Solving the first problem: AlphaFold3 on Astrocyte
Astrocyte, led by Xin Yang, is a research computing platform at UTSW that serves researchers across the institution, wrapping computational tools into a managed, web-based environment accessible without a command line. No batch scripts, no manual GPU allocation, no container configuration.
In 2025, when Google DeepMind open-sourced AlphaFold3, its breakthrough model for predicting the structure of proteins, nucleic acids, and their complexes, the Astrocyte team deployed it from open-source release to fully managed production workflow in under six weeks. This required containerizing the model within the institution's security framework, integrating downstream visualization tools so researchers could interpret results immediately, building user-facing documentation, and launching an onboarding nanocourse to drive adoption. The deployment was done entirely through manual engineering effort, without AI-assisted tooling.
The adoption signal was immediate. AlphaFold3 became the single most-run workflow on the platform within months, surpassing all other workflows including its predecessor AlphaFold2. Researchers across the institution, from biophysics to radiation oncology, Principal Investigators, and graduate trainees, adopted the tool for active research programs, with individual labs accumulating hundreds of runs as the workflow became embedded in their day-to-day research.
AlphaFold3's success on Astrocyte demonstrated something broader: that academic medical centers can move quickly to put frontier AI tools directly into the hands of researchers who need them. But the six-week sprint also exposed a structural limitation. The manual effort required to bring each new workflow to production was not scalable. Every new tool would demand the same cycle. The platform had solved the researcher's problem. It had not yet solved its own.
Solving the bigger problem: AI that accelerates the platform
This is where the story shifts from deploying AI as a workflow to deploying AI as the platform interface itself.
Next, Astrocyte platform engineer Chenlai Shi collaborated with Chinthak Murali, AI lead at BioHPC, to develop a BioHPC AI Assistant — a conversational AI layer integrated directly into the platform's front end. Instead of navigating documentation, reading parameter guides, or submitting support tickets, a researcher can now interact with Astrocyte conversationally:
"I have a set of protein sequences and I want to predict their structures and binding partners. Where do I start?"
The AI Assistant doesn't return a generic search result. It understands the platform's full range of available workflows, retrieves relevant documentation and parameter guidance, and provides a direct, actionable recommendation grounded in the institution's actual research computing environment.
This represents a fundamental shift. Historically in this context, AI was the thing being computed — a researcher submitted an AlphaFold3 job and waited for results. With the AI Assistant, AI is now also the interface through which researchers discover, configure, and launch their computations. This closes the loop between computational capability and scientific accessibility.
For platform operations, the implications go further. The same manual effort that went into the six-week AlphaFold3 deployment — building documentation, creating onboarding materials, fielding user questions about parameters and outputs — can now be partially automated and augmented through the AI layer. When the next breakthrough model arrives, the path from release to researcher adoption becomes shorter. Workflows that researchers may not have known existed become discoverable through conversation.
The result is compound acceleration. The platform already accelerated research by removing infrastructure complexity from the scientist's workflow. Now, AI accelerates the platform itself, shortening the deployment cycle for new tools while making existing capabilities more discoverable and accessible to a broader range of researchers.
IBM Fusion with content-aware storage
The BioHPC AI Assistant is powered by an integrated architecture that brings together IBM and NVIDIA technologies with the Astrocyte platform, running on Red Hat OpenShift for container orchestration.
IBM Fusion provides an integrated platform for rapidly deploying open, AI-optimized infrastructure to accelerate data pipelines, streamline IT operations, and manage inference at scale. It brings together a production-ready platform foundation, integrated data services, and the tooling teams need so they can start building, testing, and deploying AI workloads in days rather than months. Fusion simplifies operations with built-in resilience, lifecycle management, and governance controls that help keep AI environments reliable and secure as usage grows.
Fusion brings together three unique capabilities:
-
Fusion data catalog – Fusion provides the metadata, lineage and policy controls needed to ensure that stored context is discoverable, auditable and safe to reuse. It becomes the system of record for context across models and environments.
The BioHPC AI Assistant leverages Fusion’s content-aware storage capabilities, which use natural language processing to extract semantic meaning from unstructured data and keep RAG pipelines up to date and more efficient by moving document processing and vectorization closer to the data. With CAS, updates to the workflow documentation, parameter guides, troubleshooting resources, and other sources are immediately ingested and moments later reflected in the AI output, so users always get accurate, up-to-date answers.
Fusion has a search API that applications use to call CAS, which triggers the workflow to call into the vector database with a query or prompt and returns that info via the API to the AI application. Fusion has already done the vector transformation and stored the vector that represents the meaning of the underlying information in the database; it isn’t just indexing vectors into a data structure but instead is building a growing repository of the organization’s knowledge.
Leveraging NVIDIA NIM microservices and NeMo Retriever
The Fusion content-aware storage architecture relies on NVIDIA NIM, a collection of performance-optimized, portable inference microservices designed to make deploying AI models extremely fast and consistent across GPU environments. Theypackage models, runtimes, and optimizations into a single containerized unit.
Text extraction, embedding, indexing, retrieval, and reranking is handled by NVIDIA NeMo Retriever, a collection of industry-leading Nemotron RAG models. NVIDIA NeMo is an open-source AI framework for building, training, and deploying large-scale deep learning models, particularly for tasks like natural language processing. NVIDIA NIM helps optimize inferencing performance by providing pre-packaged, accelerated AI microservices that can run on NVIDIA accelerated computing across all environments, including enterprise data centers, private clouds, and hybrid infrastructures.
Storage at research scale
As GPU utilization grows with AI workload adoption, I/O between storage and compute can become the critical constraint. At research scale, high-throughput, intelligent storage infrastructure isn't optional, it's foundational.
Fusion helps eliminate storage-to-compute bottlenecks for GPU-intensive research workloads with an HCI architecture in which all the hardware components are close to each other to minimize network latency issues. It’s all racked, ready to go, with storage right next to compute. With Fusion, organizations can bring their compute resources to the storage, rather than having to copy or migrate their data.
What this means for institutions and for research
The challenges that drove this architecture at UTSW are shared across academic medical centers:
-
Non-technical user base – Unlike industry AI deployments where engineers are the primary users, researchers at academic medical centers are scientists and clinicians first. The platform must meet them where they are, which increasingly means conversational interaction, not command-line access.
-
Security and compliance – Biological data, clinical adjacency, and institutional policy create infrastructure requirements that generic cloud AI platforms cannot satisfy without significant customization. The architecture must operate within the institution's security perimeter.
-
Pace of AI model releases – The velocity of new AI capabilities relevant to biomedical research in areas such as structure prediction, molecular simulation, and genomic analysis shows no sign of slowing. The ability to rapidly onboard these tools is as important as running existing ones.
As research computing shifts toward GPU-accelerated AI workloads, the storage layer that feeds those GPUs determines whether infrastructure investments translate into research throughput. IBM Fusion is built on an architecture specifically designed to eliminate storage-to-compute bottlenecks for GPU-intensive workloads.
Today, academic medical centers are uniquely positioned to witness and shape how AI transforms life science research. The models that will accelerate drug discovery, enable precision medicine, and deepen our understanding of biological systems are arriving faster than institutions can absorb them. Platforms like Astrocyte, augmented by AI and backed by high-performance infrastructure from IBM and NVIDIA, are helping define the next era of biomedical discovery.
See It at NVIDIA GTC 2026
The Astrocyte + IBM Fusion AI architecture is being demonstrated live at NVIDIA GTC 2026 in San Jose, March 16th through 19th and showcases the end-to-end flow from a researcher's natural language query through the AI Assistant, to intelligent retrieval via IBM Fusion, to workflow recommendation and execution guidance. Visit the IBM at GTC booth #2007 to see how Fusion is helping AI-native research computing reshape biomedical discovery.