AnVIL Demos: Open Discussion Forum on March 18, 2026

AnVIL Demos
1

Topic: Open Discussion Forum

March 18, 2026 at 10:00 AM ET on Zoom

10:00 AM - 11:00 AM ET – Open Discussion Forum

In this meeting, we’ll have an open forum to chat about AnVIL, answer any questions you might have, and point you to any resources you might be looking for.

:pencil: Sign up: Register for AnVIL Demos

What are AnVIL Demos?

AnVIL Demos are a monthly, virtual meeting where we highlight what you can do on the NHGRI Analysis, Visualization, and Informatics Lab-space (AnVIL; https://anvilproject.org/), a cloud-based computing platform for genomic data science! AnVIL Demos will start out with a 30-minute demonstration on the platform followed by open time for Q&A and user support.

The demos will highlight a range of topics, from a capability of the platform to a scientific analysis powered by AnVIL. If you’re interested in showcasing how you use AnVIL at a future AnVIL Demos session, reach out to Natalie Kucher (nkucher3 / at / jhu.edu). After the demo, we’ll open up the floor to answer questions about the demo and to answer any general questions you might have about AnVIL.

:play_or_pause_button: Watch past AnVIL Demos recordings from our YouTube playlist!

Resources

Upcoming Events

Sign up to hear about future AnVIL Demos and announcements at http://bit.ly/anvil-mailing-list and learn about upcoming events at https://anvilproject.org/events!

2

Q: Interested in using AI on AnVIL. Lots of people use VSCode, this isn’t available on AnVIL. Looking at status on RStudio, could be interesting to look at terminal (cursor) or Jupyter. Does anyone have thoughts on experience or suggestions of real uses of AnVIL with a github copilot interaction?

A: In our experience, RStudio’s copilot isn’t bad, but doesn’t work well with .Rmds. Haven’t tried on HPC though.

A: It’s a really interesting opportunity, but there are risks without knowing exactly what the LLMs are doing, whether there are security gaps. It’s an ecosystem. If there are security requirements, there are permissions that need to be carefully monitored and enforced. Users need to know what they can control and what they cannot when using LLMs and unknown packages. This can be an issue if the source code isn’t validated or vetted. In research, we leverage and trust known researchers’ integrity in their code. The question is how to validate code and know it’s safe to use, like a model zoo.

Q: We share the same concerns about AI safety. From the technical side, looking into what is possible. I’m starting to play with AI copilots that autocomplete code, rather than agents. It’s a good question about policy on whether or not you can use AI on AnVIL. If a user had a package that sends out to ChatGPT, is this possible? Is there a policy blocker?

A: NIH has a policy stance on AI that users of controlled access data do not use public generative AI tools. NOT-OD-25-081: Protecting Human Genomic Data when Developing Generative Artificial Intelligence Tools and Applications

While you can use controlled access data in developing generative AI models, you cannot currently share these AI tools. So if you close out a project or dbGaP access request for controlled data, you’d have to destroy the AI model that is generated, as it’s considered a data derivative under the data use agreement. NIH is working on issuing an updated guidance. There are steps that can be taken to reduce risks of generative AI models. The guidance is to keep the data access requests open if users want to preserve the AI.

There was an RFI to the community about this topic, which had a healthy response.

Q: Are you able to use public AI to generate code on AnVIL, even if you don’t run it on controlled access data?

A: Not sure if Terra has enabled support for this. If technically it is possible, it is likely there is a user who is exploring it.

A: For data analysis, if data are structured, there aren’t many needs to approximate or detect the data organization. This could be a helpful use case for AI with new data structure, to investigate what the experiment is or what the structure is. For analysis steps, there is a lot of specification that needs to be given to run the analysis in the exact way you need it to go and a lot of review, so it can still be very hands on.

Q: It’s likely to have a broad impact to enable folks to ask questions of what is wrong with their analysis or why something isn’t working in the Terra platform.