Powering precision medicine's next frontier.
Panhuman is the biological AI that extends precision medicine across the full range of human biology, unlocking the medicine and the discoveries the field is still missing.
Technically: multimodal biomedical foundation models, trained through a federated architecture inside the institutions that hold the data. Interpretable by design. Representation as a construction requirement, not a later correction.
Precision medicine is only as good as the data it is built on.
Roughly 88 per cent of participants in the large genetic studies behind modern medicine are of European ancestry, a group that makes up about a tenth of the world's population. Continental African ancestry accounts for 0.28 per cent. Everything trained on that foundation inherits its shape.
When a test comes back uncertain for a patient in Lagos, Lima or Lucknow, that is rarely a limit of the science. It is a record of who has been measured.
Therapies and risk scores that work best for a fraction of the world, and lose accuracy where populations differ most.
Dosing, safety and interpretation that are markedly less accurate outside European ancestry. Polygenic risk scores are several times more accurate in European-ancestry individuals.
The biology in under-studied populations, and the targets it holds, remains largely unexplored.
Martin et al., Nature Genetics 2019; Sirugo, Williams & Tishkoff, Cell 2019.
The biology the field has largely overlooked.
Some of the most valuable biology in precision medicine may sit in the populations the field has studied least. Most biological foundation models are organised around a modality or a task. Panhuman is organised around representation, the axis with the most unrealised value for globally useful biological intelligence.
Why it is hard to replicate: governed, longitudinal relationships that deepen the data over time as new inputs and omics are added; an experimental feedback loop designed in from the start; and outputs built to embed in the workflows that use them.
The least-studied populations hold the most discovery.
People of African ancestry are 2.4 per cent of participants in the GWAS Catalog and account for 7 per cent of its significant associations: roughly three times the discovery value per genome. On the evidence so far, the populations measured least are among the highest-yield sources of discovery per genome, and the work of studying them properly has only begun.
It is clinically real. Of 154 known disease-causing variants examined in one recent study, 44 are common in these populations yet rare everywhere else, hidden in plain sight.
Discovery in one population, value for all. PCSK9 is now a multi-billion-dollar drug class. Its protective variants were found in African-ancestry cohorts, then became a therapy prescribed across Western markets.
The frontier has validated the category, and pointed to what comes next.
In 2025, researchers at Verily (Alphabet's life-sciences company), NVIDIA and Google published a multimodal foundation model that learns from electronic health records and genetic risk together. They trained it on 150,000 participants of the NIH All of Us Research Program, one of the world's most deliberate efforts to broaden who is represented in biomedical data. It is serious work by serious teams, and it moves the field forward.
- Multimodal models work. Clinical record and genetics learned together, in one foundation model.
- Genetics adds real predictive value. A 15.5 per cent relative improvement in type-2 diabetes onset prediction (AUPRC), with the model's own interpretability showing how genetic and clinical signals interact.
- The category is real, and the best teams are building in it.
- Representation. The genetic input was polygenic risk scores, and the authors note that scores derived from predominantly European-ancestry cohorts generalise unevenly.
- The genome itself. They identify modelling specific variants directly, rather than through summary scores, as a promising next avenue.
- More layers of biology. They call for further modalities to be brought into the model.
Those three next steps, representation across populations, whole-genome context, and multi-omic depth, are the ground Panhuman is built on. The frontier is a shared one, and we are glad to be working on it alongside efforts like these.
Amar et al., Integrating Genomics into Multimodal EHR Foundation Models, arXiv:2510.23639 (2025). Verily Life Sciences, NVIDIA and Google; NIH All of Us Research Program.
One foundation. Three ways it moves the field.
The engine: one foundation model built across the full range of human biology, and a growing family of specialised models on top, each one licensable. The signal no single-ancestry dataset can produce.
Existing therapies, everywhere they should work.
Existing therapies, applied accurately across ancestries. More patients treated well, and markets that were never properly served opened for the first time.
Today's medicine, made precise for each population.
Ancestry-aware dosing, safety and interpretation, so a result means the same thing for a patient in Nairobi as it does in Boston.
New biology from the least-studied populations.
Novel targets and mechanisms that only become visible when the model has seen the full range of human variation. Discoveries there benefit every market.
Population coverage is a design requirement, not a later correction. Training composition is recorded and reported rather than assumed.
Genomic sequence first, then proteins, metabolites and detailed clinical information, linked per individual, so the model learns how a body works rather than what one assay measured.
Pathway-guided architecture built into successive versions from the start. A prediction a scientist cannot interrogate is not usable evidence.
We are a model company. We do not hold data, we do not sell data, and we do not run clinical programmes. We build the intelligence layer and license it.
The data never moves.
Panhuman trains through a federated architecture. Analysis runs inside each contributing institution's own environment, on infrastructure it controls. Only model parameters are exchanged. The underlying data is never transferred to us, and we never hold it.
This is not a compliance workaround. It is the only architecture under which institutions in most of the world can participate at all, and it is why the model can reach populations that centralised approaches cannot.
Training runs where the data already sits, under the institution's control, on infrastructure it operates.
Model updates are exchanged. Records are not. Nothing that could identify a participant leaves the site.
Assay type, capture method, sequencing depth and batch conditions are modelled explicitly, so differences between cohorts are not mistaken for biology.
Each contribution strengthens the shared model. Each contributor receives its own node model, built for its population, at no cost.
Every new genome strengthens every future insight.
A model trained on one population understands that population. A model trained across many learns which signals travel and which do not, and that is a different and more valuable thing.
The advantage compounds. Broader representation makes the model more useful, which makes it more attractive to contribute to, which broadens representation further. Compute is cheap and base architectures are open. The scarce input is reach, and reach is earned rather than bought.
Each contributor receives a node model built for its own population, so an institution that joins does not only strengthen the shared model. It gains a capability of its own, one it did not have before.
What exists today, and what is coming.
The base model already reflects breadth no single-ancestry dataset can produce. The current partner pipeline carries the potential to improve representation for populations totalling 3.05 billion people: an increase at the scale of civilisations, not cohorts.
Cohort-specific, governed data access is developed in phases with participating institutions. Signed agreements validate the governed contribution model, and each new region joins a network that returns more than it asks.
How contributors benefit: the flywheel, no-cost local model access, and a regional share of value.
Every scientific milestone is a product.
Publishable outputs at every version, per population, co-produced with local scientific investigators. Each is publishable proof and a licensable product, and the relationships they build are longitudinal and governed.
- Science
- Allele frequency, haplotypes, population pharmacogenomics
- Product
- Ancestry-, age- and sex-specific dosing
- Science
- Variant reclassification, population reference ranges
- Product
- Resolves "uncertain" results in non-European patients
- Science
- Protein expression and pathway signal across populations, protective biology, target validation
- Product
- Novel druggable targets: the enzymes, receptors and other proteins that medicines act on
- Science
- Transcriptome, metabolome and further layers integrated, predictive and generative
- Product
- Editing-ready target design
Three commitments shape how the models are built. Each costs something, and each is why the result can be trusted.
Representation is a construction requirement
Population coverage cannot be corrected after training. A model's representations are formed by what it saw, and a supervised layer added later maps those representations to labels without changing them. Broadening who a model understands is a sequence-level operation, and it has to happen during training.
We record training composition and report it, together with a coverage index that shows where the evidence is thin. Where a question falls outside what the training data supports, the model says so rather than returning a number that looks like an answer.
Mixed data quality, handled explicitly
Contributing cohorts differ. Whole genomes, exomes and targeted panels carry different information and support different claims. Sequencing platforms, capture methods, depth and library preparation differ between sites, and those differences correlate with geography and therefore with ancestry.
Treating them as interchangeable would produce ancestry-linked findings that are artefacts of laboratory practice. Every sample carries its assay characteristics as explicit covariates, claim scope is tracked per output, and results stratified by population are controlled for them.
Interpretability, built rather than retrofitted
Interpretable-by-design, pathway-guided architecture is a requirement of successive model versions from the outset. In target discovery, a prediction that cannot be traced to a mechanism is a lead nobody can act on, and in regulated settings it is not evidence at all.
Governance is infrastructure, not overhead.
An institution that cannot leave a programme was never really a participant in it.
Institutions in underrepresented regions have been asked for their data before, often on terms that returned little to them. Their caution is well founded. We build on a governance model designed to earn that trust and to be verifiable independently of anything we say.
Contributing institutions do not sign with Panhuman. They accede to the charter of the Count Us In Foundation, an independent non-profit that admits participants, approves how data may be used and distributes the benefits that arise. Under that structure, institutions retain sovereign ownership of their data at all times, hold seats on the councils that govern the programme in their region, and receive a share of what the programme generates. Participation is a decision they make and can reverse.
Panhuman's role is bounded, and deliberately so. We develop the models. We do not receive, hold or process participant data. We hold no seat on the Foundation's councils, and we are not the counterparty institutions negotiate with.
If you hold data, steward a population, build models, or need them to work everywhere.
Contribute without your data leaving your environment.
A governed programme that returns model access, capacity and a share of what your contribution creates, with your cohort's data never leaving your control.
A model for your population, inside your borders.
Build a national node: training on your own infrastructure, under your own jurisdiction, contributing to a shared model and holding a node model of your own. Precision medicine capacity that stays in the country.
Targets validated where your evidence base ends.
Reach, refinement and discovery across populations your existing evidence does not cover, drawing on some of the highest-yield discovery populations per genome in the field. Licensed, recurring.
An unsolved engineering problem at global scale.
Federated infrastructure, harmonisation and evaluation across dozens of jurisdictions. We are interested in people who find that appealing.
Panhuman Inc.
Panhuman Inc. is a Delaware corporation founded in 2026. We are an AI research company building multimodal biomedical foundation models trained on globally representative human data.
We were established to carry a foundation model programme forward in a dedicated entity, so that model development sits with a party that holds no data and has no national affiliation. That neutrality is a design decision. It is what allows institutions and national programmes in different jurisdictions to contribute to one shared model without contributing to one another's.
Our first substrate model is developed. Outputs are produced per population and co-authored with local scientific investigators, so that each result is publishable proof as well as a licensable product.
The vision: a model that reflects humanity, not subsets of it, where every new genome strengthens every future insight.
Panhuman Inc., Delaware, United States · contact@panhuman.ai