Why We're Open Sourcing Our First Model

A note from Jenny Yang, Co-founder & CEO of Outpost Bio

Progress is a team sport

Before Outpost, I spent years in research, and the common thread across all of it was collaboration.

At the Genome Sciences Center in Vancouver, I was part of a team working on personalized oncology. Clinicians, bioinformaticians, and ML practitioners were all pulling on the same problem from different angles. When we were looking for specific genomic markers to guide treatment, one of the first questions was always: are there patients elsewhere in Canada we can learn from? A single center's data was never enough. The signal was in the aggregation, across cities and institutions, because human health isn't one-size-fits-all and you can't truly understand treatment outcomes from a single slice of the population.

At Oxford, I worked on COVID-19 triage models for emergency departments. Those models performed as well as they did because four hospitals shared their data, because researchers and clinicians worked together to figure out what the models actually needed to do, and because when we brought the work to Vietnam, clinicians there shaped it for their reality at the bedside. Building AI that actually works in clinical settings takes so many different kinds of expertise, and the real magic is in getting those perspectives to come together.

The technical side stood on open shoulders too. Whether I was building novel architectures across Oxford and Exscientia or pushing into new problems, I was almost always starting from someone else's paper, someone else's open-source repo, someone else's willingness to put their work out before it felt "finished." ML research moves at the pace it does because of that culture, and I've benefited from it at every stage of my career.

We wouldn't be here if the people before us hadn't shared. I'm not willing to pull that ladder up behind me, and I don't want anyone to think I climbed it alone.

The microbiome is too big for any one lab

The microbiome is one of the hardest frontiers in biology. On the wet lab side: anaerobic organisms, fragile communities, and protocols that behave differently between labs. On the dry lab side: massive, noisy, high-dimensional data where reproducibility is still an open problem. And the scope is enormous: trillions of organisms, thousands of species, functional redundancy everywhere you look.

Everyone agrees the microbiome matters for health and disease. The community hasn't cracked how to decode it reliably yet, and that's exactly why this can't be solved behind closed doors. It needs peer review, reproducibility, and a lot of people pulling in the same direction. Putting this model out open source is how we invite the rest of the field in: dig into it, plug it into your pipelines, tell us where it breaks. The faster we share what we're each learning, the faster any of us can make real progress for patients, for the researchers and clinicians working alongside them, and for anyone invested in understanding human health.

AI for health has to be built with the people it's for

I care a lot about responsible AI. It was the subject of my PhD, and it's still the thing I think about most when we build. A big part of that, for me, is making sure the people who will actually use this get to shape it from day one. Researchers need models that slot into their existing workflows. Clinicians need outputs they can act on without a PhD in ML. Drug and food companies need insights specific enough to inform real product decisions. And the broader public deserves a seat at the table, because the whole point is to give more people more agency over their own health.

Open sourcing our first model is the most straightforward way I know to start that conversation in good faith. Here's our work, here's how it was built, here's what it can and can't do. Use it, critique it, make it better. We want to build this for the world, and with it.

A foundation, not a finish line

I want to be clear about what we're not claiming. We haven't solved the microbiome. Biology is too complex, and every use case, whether it's a specific disease, population, drug, or ingredient, will need its own treatment. A foundation model is exactly what the name suggests: a starting point, not an answer.

What we're offering is something strong to build from. Fine-tune it to your question. Pair it with your data. Adapt it to the context you actually care about. My hope is that because you started with our model, you're not starting from scratch.

That's also how we'll keep building at Outpost. This foundation sits underneath everything, including the commercial models we'll develop for specific use cases. Those will be built with proprietary data, but the base they stand on will always be open. And where we can contribute data back publicly, we will.

The same instinct that pushes me away from "one size fits all" medicine pushes me toward this kind of release. Personalized health means different people, diseases, and contexts need different tools. Responsible AI means giving people the ability to take the right steps for their specific purpose, not handing them a black box and calling it universal. Open sourcing the foundation is how we make that possible for the people best positioned to build on it.

- Jenny