Simplifying AI drug discovery on HPC

REANNZ's technical expertise is making advanced AI drug discovery tools easier for University of Auckland cancer researchers to access and use.

Research background

NVIDIA invokes a ‘flywheel’ concept to describe its GPU hardware/software integration, with a self-reinforcing cycle of faster hardware, more ambitious models and increased demand for better tooling, which in turn justifies further hardware investment. In drug discovery, this flywheel has compressed timelines that once measured in decades into workflows that run in hours.

MolMIM was an early expression of this approach in molecular generation — a generative AI model that helps researchers generate molecules with desired properties by navigating the learned internal representation of chemical space, guided by a user-specified scoring function. It has since been succeeded by GenMol, a discrete diffusion-based model using the Sequential Attachment-based Fragment Embedding representation, whose parallel decoding architecture makes it more efficient and versatile across complex, multi-objective drug discovery workflows.

Where these tools generate and optimise candidate molecules, Boltz-2 evaluates them against biological targets. Built as a foundation model for biomolecular interactions, Boltz-2 predicts binding affinity — how tightly small molecules attach to proteins — overcoming the longstanding trade-off between the accuracy of expensive atomistic simulations and the speed of less precise docking methods.

The image output by NVIDIA's MolMIM tool.

Figure 1: Molecules designed by MolMIM, ranked by drug-likeness. Each panel shows a molecule generated by MolMIM, displayed as a structural diagram where lines represent chemical bonds and letters represent atoms. The number beneath each molecule is its drug-likeness score — how closely it resembles the kinds of molecules that make good medicines — on a scale from 0 to 1. The molecules are arranged from most drug-like at the top to least drug-like at the bottom. All of the molecules shown score above 0.8, meaning that MolMIM has successfully generated a set of strong drug candidates in a matter of seconds. https://build.nvidia.com/nvidia/molmim-generate

Project challenges

For researchers, the promise of NVIDIA's drug discovery stack sits behind a formidable operational barrier.

Each tool — MolMIM, GenMol, Boltz-2 — is distributed as a NIM (NVIDIA inference microservice) requiring NGC API credentials, container runtimes, and GPU infrastructure that most academic environments were not designed to provision. A researcher wanting to run a single binding affinity prediction must first navigate API key registration, pull multi-gigabyte container images, configure GPU drivers compatible with local CUDA drivers and manage service endpoints that assume persistent cloud infrastructure rather than the job-scheduler models common to HPC clusters.

Interactive API access compounds this further — workflows designed around notebook-style iteration clash with batch compute environments where sessions are time-limited and outbound network access restricted. The result is that the scientists best positioned to use these tools spend too much time on infrastructure rather than discovery, precisely inverting the productivity gains the flywheel was designed to deliver.

What was done

For University of Auckland cancer pharmacologists Shanan Chand and Stephen Jamieson, access to these tools with minimal overhead is imperative. They contacted REANNZ's Research Software Engineering group to develop solutions that simplified NGC API key access, converted NVIDIA's container format to apptainer format for compatibility with REANNZ's GPU infrastructure, and provided a streamlined entry point to each model's API through the REANNZ Open OnDemand JupyterLab interface.

Because NVIDIA requires access keys across more than 100 models in their API catalogue, REANNZ engineers designed a module that checks whether a user has access to a particular model, and if not, directs them to the relevant access request page before securely saving the key for future use. This provides a minimal but instructive introduction to NVIDIA API key management.

Container conversion was implemented as a just-in-time process from Docker, ensuring model functionality remains consistent across different platforms regardless of the underlying container runtime environment. Finally, Open OnDemand gives users a simple graphical user interface to select which REANNZ GPU to use, while JupyterLab's Python environment allows API requests to be expressed as reusable objects rather than raw calls inside a programming environment.

The result is that software development can now proceed with the complexity of NVIDIA's container, GPU, and API stack abstracted away — placing the focus back on the science.

An output from NVIDIA's tool.

Figure 2: A protein-drug interaction predicted by Boltz-2, coloured by confidence. The ribbon diagram shows the three-dimensional folded structure of a protein, with a small drug-like molecule (shown as a multicoloured stick figure) nestled in its binding site near the bottom. The colour of each part of the protein reflects how confident Boltz-2 is in its predicted shape — yellow and green regions are predicted with high confidence, while teal and purple regions are less certain. The header reports two key results: an IC50 of 8.45, which is a measure of how potently the molecule is expected to inhibit the protein's activity, and a binding affinity likelihood of 0.93 out of 1, indicating that Boltz-2 is highly confident this molecule will bind to this protein. Together these numbers suggest the molecule is a strong candidate for further investigation as a potential drug. https://build.nvidia.com/mit/boltz2

Researcher feedback

"The consultancy we received from REANNZ was crucial to allow us to access and implement molecule generation packages to fast-track our drug discovery projects by evaluating a vast range of chemical space to identify new drug compounds."

- Shanan Chand, Auckland Bioengineering Institute, University of Auckland

 


 

This case study shares some of the technical details and outcomes provided through our Consultancy Service. This service supports projects across a range of domains, with an aim to lift researchers’ productivity, efficiency, and skills in research computing. Get in touch to discuss how our Research Software Engineers and specialist support could help advance your project.

Find anything about our products, services, and more. Enter a query in the search input above.