A kidney can be described in sequence, in cells, in tissue and in a clinical course, and each description on its own is partial. We train a model for each and hold them in one representation, so that a target can be traced from a genetic variant to a patient outcome and then made into a medicine. The platform is the means. The therapy is the point.
General-purpose cell models are trained on every tissue and know the kidney only in passing. Ours is trained on the organ, so it resolves states that generic embeddings blur together.
Every nephrology patient who is biopsied already has a slide. NephroLens turns that slide into numbers rather than a grade, and into molecular structure rather than morphology alone.
Kidney function is followed with a number that moves slowly and late. This model learns the shape of the course itself, from records and blood measurements, and gives a call earlier than the number does.
Most variants associated with kidney function sit outside genes. This model predicts their regulatory consequences from sequence and anchors the prediction in kidney chromatin.
A snapshot is not a twin. What makes it one is the update loop: serial samples, treatment and outcome, fed back so the model is corrected by what actually happened.
The variants a patient carries, and what those variants do to regulation in the kidney.
Which cell types are present, in which states, and how they are talking to each other.
Where those cells sit in the section, and how the architecture of the organ is changing.
Function, treatment and outcome, resynchronised with the biology at every visit.
Each of these papers unlocked a capability the platform now depends on. They are peer-reviewed and can be checked. What is proprietary is the trained models and the programmes they point to.
Gave us a ranked, causal shortlist of genes rather than a list of associations.
Connected genotype to protein to disease, which is what makes a blood readout possible.
Showed where each cell sits and what it does to its neighbours, the ground truth the models learn.
Separates states that general-purpose cell models blur, which is the basis of NephroBase Cell.