Astellas Pharma used NVIDIA’s Boltz-2 NIM to accelerate membrane-protein drug discovery, cutting physical assays by 90% and research time by 70%, the company reported this week. The AI-guided approach correctly identified 10 active compounds from a pool of computationally screened candidates—an unprecedented hit rate for a notoriously difficult target class. The same AI microservice is now accessible on Windows workstations and servers, bringing high-performance computational drug design to a far wider pool of researchers.

What just happened: a 90% slash in physical screening

Membrane proteins sit at the surface of cells and control communication, nutrient transport, and drug resistance. They’re also among the hardest targets to drug—too flexible, too hydrophobic, and too difficult to isolate for traditional high-throughput screening. Astellas, one of Japan’s largest pharma companies, turned to NVIDIA’s Boltz-2 NIM to short-circuit the usual trial-and-error cycle.

Boltz-2 is a generative AI model trained to predict the three-dimensional structures of protein-ligand complexes. It takes a protein target sequence and a chemical library as inputs and ranks molecules by their likelihood of binding. Astellas deployed the model as a NVIDIA Inference Microservice (NIM) on Tokyo-1, an NVIDIA DGX-based supercomputer operated by Xeureka (an Astellas subsidiary). The system ran structure predictions at scale, filtering a large compound library down to a handful of top candidates.

When those candidates were synthesized and tested in the lab, all 10 showed the expected biological activity. The traditional approach—running thousands of physical assays to find a single hit—would have required 10 times the bench work and months of additional time, Astellas told NVIDIA. “The integration of Boltz-2 NIM into our workflow has fundamentally changed the speed and accuracy of our early-stage discovery,” the company said in a statement.

The numbers are stark: 90% fewer physical assays, 70% shorter research cycles. For membrane-protein targets, where success rates often languish in the single digits, a 10-for-10 hit rate rewrites the risk calculus of early-stage research.

What it means for you—whether you’re a researcher, IT admin, or just a Windows user

If you’re a researcher in biotech or pharma, this announcement is more than a vendor case study. It’s a signal that cutting-edge AI—the same microservice Astellas used—is no longer locked inside a supercomputer. NVIDIA AI Enterprise ships Boltz-2 and other BioNeMo models as containerized NIMs that can run on Windows 10/11 workstations and Windows Server 2019/2022. All you need is a compatible NVIDIA GPU (an RTX 3080 or better for molecule-scale work, or an A-series data-center card for production pipelines), the latest Windows drivers, and a Docker or WSL 2 runtime.

For IT administrators in life-sciences organizations, the practical upshot is that you can now deploy the exact same AI inference stack that powered the Astellas breakthrough on your existing Windows fleet—no migration to Linux required. NVIDIA provides pre-built containers with model weights and optimized CUDA kernels, plus enterprise support and patching. You can manage access via Active Directory, run jobs on local GPU workstations during off-hours, or burst to Azure’s NCasT4_v3 instances (which use the same NVIDIA GPUs) when a research team needs more compute.

Everyday Windows users might wonder what any of this has to do with them. The indirect answer: faster drug discovery means potential new therapies for diseases that have stymied traditional chemistry. Membrane proteins are implicated in cancer, autoimmune disorders, and antibiotic resistance. When a pharma company can screen thousands of virtual compounds in days instead of months, the pipeline of drug candidates enters clinical trials faster—and at a fraction of the cost. That doesn’t lower your prescription bill today, but it reshapes the economics of an industry that often spends 10–15 years and billions of dollars per approved drug.

There’s also a direct angle for students, citizen scientists, and startup founders. NVIDIA’s NIM catalog includes free, rate-limited access to smaller BioNeMo models. A grad student running Windows on a gaming laptop with an RTX 3060 can prototype a ligand-screening pipeline using the same APIs that Astellas paid for. The gap between “hobbyist” and “enterprise” AI has rarely been narrower.

How we got here: from AlphaFold to a turnkey Windows NIM

The story starts in 2021, when DeepMind’s AlphaFold2 cracked the protein-structure prediction problem and open-sourced its model. Suddenly, researchers could predict the shape of a protein from its amino-acid sequence with near-experimental accuracy. But structure alone isn’t enough to design a drug; you need to know which molecules will bind, and with what affinity. Predicting protein-ligand interactions—the holy grail of computer-aided drug design—remained a stubbornly hard problem.

NVIDIA, which had been building its Clara healthcare platform since 2018, saw an opportunity. In 2022, it released BioNeMo, a framework that let researchers fine-tune large language models on molecular data. The early BioNeMo models could generate novel molecules and predict properties, but they weren’t yet integrated into a single-click inference service. The breakthrough, announced at GTC 2024, was Boltz-2: a diffusion-based generative model that directly predicts the full 3D complex of a protein and its ligand. By wrapping it in a NIM—a secure, optimized container that exposes a standard API—NVIDIA made the model as easy to deploy as a web server.

Astellas, already a BioNeMo partner, became the first big name to publicly validate Boltz-2 at production scale. Tokyo-1, its purpose-built NVIDIA supercomputer, gave the company a home-field advantage. But the architecture of a NIM is the same whether it’s running on a thousand H100s in a data center or a single RTX 4090 under a desk. That architectural sameness is what lets Windows users to participate.

Microsoft and NVIDIA have been quietly tightening Windows support for AI workloads. Windows Server 2022 added GPU partitioning and DDA (Discrete Device Assignment) features aimed at virtualized AI inference. Windows 11 24H2 extended the Windows Copilot Runtime—not directly relevant to pharma, but a sign of Redmond’s willingness to treat GPUs as first-class resources. The NVIDIA AI Enterprise license, which covers NIMs, is available through the Azure Marketplace and certified for Windows. As of March 2025, the installation docs explicitly list Windows 10 (21H2 and later) and Windows Server 2019/2022 as supported hosts.

What to do now: a practical checklist

If you want to replicate Astellas’s workflow—or just test the waters—here’s how to get started on a Windows machine.

  1. Check your hardware. Any NVIDIA GPU with at least 16 GB of VRAM will run small molecule docking with BioNeMo. For the full Boltz-2 model, aim for 24 GB or more (RTX 3090, 4090, or A-series). The NIM download page lists exact requirements.
  2. Set up the runtime. Install the latest Game Ready or Studio driver from NVIDIA’s website, then enable either Docker Desktop with WSL 2 backend or hook your container runtime directly to WSL 2. NVIDIA’s container toolkit is available as a Windows-native installer.
  3. Get a license. NVIDIA AI Enterprise is free for 90 days for evaluation; after that, subscriptions are per-GPU/per-year. Academic institutions get discounts. The BioNeMo NIMs, including Boltz-2, are included in the "BioNeMo Early Access" section of the NVIDIA NGC catalog—you’ll need to sign in and accept the terms, then generate an API key.
  4. Pull and run. With the NGC CLI installed, a single command pulls the Boltz-2 NIM container: ngc registry model download nvcr.io/nvidia/clara/bionemo_boltz-2_nim:latest. Then launch it with docker run specifying your GPU. The container exposes a REST endpoint at localhost:8000—you can submit protein sequences and SMILES strings via cURL or a Python script.
  5. Validate. NVIDIA ships a sample dataset with the container. Run the provided inference notebook (Jupyter Lab is bundled) to predict the binding pose of a known drug like imatinib on its kinase target. If the output matches the crystal structure, your stack works.

For IT teams, there are extra steps: configure Active Directory authentication for the NIM APIs, set up SSL certificates, and integrate the endpoints with internal workflow tools like KNIME or Pipeline Pilot (both have Windows clients). Microsoft’s forthcoming Windows AI Studio in Visual Studio Code might eventually offer one-click deployment templates for these scientific NIMs, but for now the CLI path is the only official route on Windows.

What we’re watching next

Astellas’s results are a single data point—impressive, but not yet a replicated clinical pipeline. The next milestones to watch: peer-reviewed publications from other BioNeMo adopters, integration of Boltz-2 predictions with cryo-EM reconstruction (hinted at in NVIDIA’s GTC 2025 healthcare keynote), and—critically—the first Investigational New Drug (IND) application that cites an AI-chosen lead compound as the principal candidate. On the Windows side, keep an eye on Build 2025. Microsoft has previewed a “Windows AI Platform” that could bundle containerized NIM support into the OS, removing the friction of the current Docker-plus-CLI setup.

In the near term, the message is simple: the same generative AI that upended public conversation with ChatGPT is now upending drug discovery, and it’s arrival on Windows means that anyone with a capable GPU—not just a supercomputer center—can join the search for the next generation of life-saving molecules.