How the Garvan Institute is changing the way it does science with Claude

The Garvan Institute of Medical Research is one of Australia's leading biomedical research institutes, with deep expertise in genomics, immunology, and cancer. In March 2026, Anthropic announced a collaboration with Garvan, providing a significant investment in Claude credits to support the institute's genomics research. The institute has adopted Claude across its research and operations, and two of its senior researchers are recipients of support through Anthropic's AI for Science program: Professor Joseph Powell and Professor Daniel MacArthur (jointly based at Garvan and the Murdoch Children’s Research Institute). We spoke with Garvan's Chief Scientific Officer Professor Sarah Kummerfeld, along with Powell and MacArthur, about upcoming research breakthroughs in drug discovery and rare disease diagnosis, and how adopting Claude across the institute is changing the way their teams approach science and discovery. 

Anthropic: Joseph, your team has been building one of the largest genomic data resources in the world. What are you working toward?

Joseph Powell, Garvan Institute: Our research seeks to address one key challenge: how the genetic differences between people functionally act to cause disease. Once we learn that, we can use it to guide the right therapies to patients and, most importantly, accelerate how we identify new targets for drug development.

We have enormous data resources where we've generated genomic information on individual cells from tens of thousands of people. It's hundreds of terabytes of data. We're using AI tools to accelerate the way we analyse that data to answer those questions. We have a series of major work coming out soon, and we're at an extremely exciting inflection point. Through the analysis of our data, we've been able to identify the specific mechanisms by which about 40% of all disease genetic factors act at the level of individual cells. In some cases, there are already treatments on the market that address some of those mechanisms. But in the vast majority of cases, there are not.

That's a striking claim. 40% of disease risk traced to specific cellular mechanisms. What role is Claude playing there, and what does it mean for patients?

Powell: We're using Claude to accelerate the way we can take those findings, identify and inform experiments, and design new treatments for the targets we've discovered. It's a dramatic speed-up in the time it takes to go from a genetic signal to a potential therapy. The acceleration we've had through the intersection of genomics and AI has gotten us to the point where I am confident that by the time that my son—who's currently in primary school—is an adult, we will have effective treatments available for the vast majority of diseases that affect society today.

Daniel, your work focuses on getting a diagnosis to families affected by rare genetic conditions. How big is the gap right now?

Daniel MacArthur, Centre for Population Genomics (a joint initiative of the Garvan Institute and Murdoch Children’s Research Institute): Our Centre focuses on ways genomics can improve medicine, and in particular how these technologies can be implemented in a way that's equitable across different populations. Our projects include a partnership with underrepresented communities across Australia to make sure they are represented in future genomic resources. We also have a really strong focus on improving the diagnosis of genetic disorders, which are individually rare but collectively affect about one in every 17 Australians.

These are diseases like muscular dystrophy or cystic fibrosis, each of them caused by just one or two genetic changes somewhere in an individual’s genome. The process of genetic diagnosis means finding those one or two changes out of the millions of DNA differences present in every person’s genome. Genomic technologies make that process possible, but right now, we are unable to give over half of all the families affected a clear answer as to exactly what DNA change is responsible. That has real implications: it means they are less likely to get access to family counselling, less likely to understand what the future looks like for their child, and much less likely to get access to targeted therapies or clinical trials.

What's standing between those families and a diagnosis?

MacArthur: It's not sequencing. It's now straightforward to generate a sequence of all 3 billion letters of DNA. The bottleneck is that the standard process for interpretation of genetic variants identified in DNA sequence information is dependent on many, many hours of expert human labor by a small group of experts: pulling together information from databases, clinical registries, published literature, and computational predictions. That can take many hours per variant, and multiple variants may be identified for each patient. It just doesn't scale.

And there's a compounding problem. If your first test doesn't give an answer, you may wait many years before it's analyzed again, even though new disease genes are being published every week. The answer to your child's condition might have been sitting there for the last couple of years, but no one has had the time or capacity to look. We're building AI tools to change that, so we can give answers back to families not in years or even weeks, but ideally in hours.

How are you using Claude specifically to tackle that?

MacArthur: Our team is working with Anthropic, as well as researchers and engineers at the Murdoch Children’s Research Institute and the Victorian Clinical Genetics Service, on building a multi-agent approach to make this process much faster. These are teams of AI agents, each specialised to search for a particular type of information relevant to a patient's diagnosis. One might check population databases. Another searches clinical registries. Another monitors the published literature. What we’re building towards is teams of agents that work together 24 hours a day, constantly looking at the latest information and updating human curators any time there's something that could be relevant to a particular child's case.

The power of tools like Claude is that they allow us to take all of these different complex types of information and weave them together in ways that human experts can instantly make use of, and that provide easy approaches to validate against the evidence. The goal here isn’t to replace expert human labor, but to allow clinical curators to make more sense of more data at the right time. We think the model of AI tools empowering human expertise is the only way to make this work at scale.

Sarah, you're overseeing the institute's broader adoption of Claude. How is it changing the way research gets done?

Sarah Kummerfeld, Garvan Institute: Claude is really changing the way we think about the work that we do. From the most basic tasks to large-scale analyses that people previously wouldn't have been able to do themselves. And people who are already technical are so much more efficient, so they can think about problems at a higher level rather than getting down in the weeds.

You've described your PhD students as "becoming managers." What do you mean by that?

Kummerfeld: Typically as a PhD student, you're doing all the work, and writing every line of code, and it takes a long time. With Claude, even our PhD students are essentially becoming managers. They're managing a team of agents, which means they can be a lot more productive, think at a high level, get into the science, and come up with hypotheses in a way that wasn't possible before.

It's also making research more interdisciplinary. Research becomes very focused very quickly. People are working on a single gene, trying to come up with a way of solving a disease that impacts that one gene. Claude lets us have an individual person interrogate knowledge from all of those different fields, including areas well out of the researcher’s wheelhouse. That wasn't possible when you had to find a human expert in every single one of those fields.

What does this mean for drug development specifically?

Kummerfeld: If you look at developing a new drug, the percentage of drugs that get through every stage and actually get to market is tiny compared to the number in research and development. AI means you can do a better job of screening your hypotheses and make sure you're not spending a huge amount of time in the lab on things that are not going anywhere. You can integrate so many different data sets and bring them together in ways that weren’t possible before.

We've been using machine learning in this space for a long time. Twenty-five years ago, when I was working on protein structure prediction, we were using machine learning and thought we were doing pretty well, getting about 60% accuracy. Now, the models are so much better that you can actually predict protein structure using machine learning without doing an experiment. When you're thinking about designing a new drug, that whole concept of rational drug design is actually becoming a reality. Large language models can help you identify a target and then protein structure prediction tools allow researchers to design new antibodies. Where we had to do a lot of trial and error in the lab, we can now come in with a very specific hypothesis.

There's a common concern that AI will shrink research teams. What are you seeing?

Kummerfeld: I don't see our workforce shrinking. I think the nature of the work will change. Menial tasks can be handled by the agents, and those roles will shift into much more high-level thinking roles. 

Daniel, what would you tell other research organizations thinking about adopting AI?

MacArthur: Claude Code has completely transformed the way that I work as a scientist and as a leader. I’ve used it to build new data visualization portals, to make sense of complex data sets, and to help me learn new topics and bring together information to make better decisions. 

I've spoken to a lot of other research leaders who say they're just too busy to try it out, but they’re excited about their postdocs using it. My response is that research leaders are actually the people who can benefit the most from these technologies: they already have deep experience in carefully designing research questions and interpreting complex results, exactly what’s needed when working with agent-based approaches. It’s hard to imagine a higher-leverage activity right now for a lab head than spending a few afternoons getting up to speed with a system like Claude Code. 

For organizations, it’s easy to express excitement and support for AI, but the big test is how fast you can go from having a postdoc who is excited about using agentic coding tools, to the point where they're actually using those tools to write real software. It’s not just about setting up an enterprise agreement. It's about figuring out how you can support your teams, build the right guardrails, and get everything cleared out of their way so they can start executing as quickly as possible.

It’s worth noting that this isn’t easy, and may require substantial changes to things like IT governance. But I expect we will see a massive difference over the next couple of years in the output from research organizations that successfully empower their teams to use agentic tools.