Lead Machine Learning Scientist, Genomic Foundation Models

The problem
Almost every AI model now being built in healthcare inherits the shape of the data behind it, and that data describes a narrow slice of humanity. 88.4% of participants in genome-wide association studies are of European ancestry (GWAS Diversity Monitor, University of Oxford). Most of the world's people are close to a rounding error in the evidence base that the next generation of biomedical AI is being built on.
Panhuman builds foundation models of human biology that include the populations medical research has left out. Models are trained inside the countries and institutions that hold the data, rather than by moving the data, and what leaves is auditable by the institution before it goes.
The role
You would own the model half of the programme, working alongside our CTO and reporting to the founder.
That means the training and adaptation architecture, the evaluation harness, the discipline of claiming only what the evidence supports, and being the person who answers when a national genomics programme's head of AI asks how it actually works.
This is an early-stage company. The scope is unusually wide, the work is visible, and there is no team to manage yet.
What we are looking for
You have trained or adapted a biological, genomic or protein foundation model yourself, rather than used one. We will want to see it: a named model, a preprint, a repository, a model card where your contribution is identifiable.
You have done parameter-efficient fine-tuning of large sequence models in practice, and can say what you chose and why.
You have run distributed multi-GPU training as an operator. You know what breaks at scale and how to recover from it.
You know what deep mutational scanning, MPRA, eQTL and ClinVar are, and specifically why clinical label databases are a biased target to evaluate against.
You can think clearly about memorisation and leakage, because our counterparties ask and the answer cannot be improvised.
Helpful, not required
Population or statistical genetics. Federated or privacy-preserving training. A publication record. Prior work inside a regulated or sovereign data environment where models and outputs could not freely leave.
Practical
Remote. We are open to candidates anywhere, and we have a mild preference for timezones outside North America, with a few hours of working overlap with Europe and the Gulf. We can employ or contract across several jurisdictions and will work out the right structure with the right person. Full time preferred; we will consider a contract arrangement.
Where we will be
Our founder will be in Toronto on 16 October, at ASHG in Montreal on 23 and 24 October, and in New York on 25 October. If you are in or near any of those cities and this is your field, we would rather meet in person than exchange emails. Apply and tell us which city works for you.

Work arrangement
Yes