Position statement

What we are trying to do.

We are building an AI system to hold the whole state of knowledge in a rare disease at once — to find the connections and relationships across conditions that no single specialist is positioned to see, to surface drug-repurposing candidates and therapeutic hypotheses worth testing, and to hand each one to researchers with the experiment that would disprove it.

We are beginning with CASK.

Why this field

Where a small amount of help goes furthest

Rare disease is under-studied not because the science is uninteresting, but because the economics are unforgiving. That is exactly what makes it the right place to point this work.

The scale

Ten thousand conditions, almost no treatments

The NIH’s Genetic and Rare Diseases Information Center counts more than an estimated 10,000 rare diseases, and states that fewer than 5% have approved treatments. Individually each is small. Together they are not.

The asymmetry

A handful of papers, one or two labs

A common disease has thousands of researchers, a deep literature, and sustained industry investment. A rare one may have a few dozen papers, one or two labs that know it, and a family foundation trying to fund the next experiment out of what it can raise this year.

The leverage

The scarce resource is attention

Where thousands of people already read the same literature, one more careful synthesis changes very little. Where three do, it can change what gets run next. The same effort buys far more here — and that asymmetry is the argument for doing this work in rare disease rather than somewhere better resourced.

So the aim is leverage: to help a small budget’s next grant, or a single lab’s next experiment, land on the thing most worth doing. We cannot add researchers to an under-resourced field. We can try to make sure the ones already in it are not spending a scarce year on a question the existing literature could have answered, or missing a connection to a better-studied condition because nobody had read across both.

In practice

Four things, in order

Each depends on the one before it. The reason the first is unglamorous is that the fourth is worthless without it.

Read everything, and know what it is worth

A rare disease has a small literature of wildly varying quality, and one weak paper repeated often enough becomes folklore. We read it all and keep a calibrated judgment of how much each finding can bear — because a synthesis built on unweighted evidence is confident noise.

Find the connections across conditions

Mechanisms do not respect disease boundaries. A pathway implicated in one rare condition often appears in several, and what has been measured in a well-studied gene may bear on one nobody has studied. This is where an ultra-rare disease can borrow statistical power it will never have on its own — and it is the part a machine can genuinely do better than a person, because nobody reads across thousands of genes at once.

Look for something that already exists

A compound already studied for a condition that shares a mechanism is a far cheaper and faster starting point than a new molecule. Repurposing candidates are the first place we look, and a shared mechanism is a reason to investigate — never a reason to treat.

Hand over something a lab can kill

The output is a falsifiable hypothesis: a specific claim, its warrant, and the single cheapest experiment that would settle it. We rank candidates by how cheaply they can be disproved, not by how promising they sound. That is what makes it safe to act on a generated hypothesis at all.

What we drew ourselves

Before anything can be connected, someone has to read the papers and say what they claim. These are the relationships the panel extracted from the literature itself — each one attached to the paper that made it, and to the reasoning that produced it.

10,421 relations drawn from 764 papers by the panel that read them. Unlike the licensed substrate shown next, these are our own work product rather than a licensed database — which is why they can be shown by kind rather than only by count. A relation here is a claim a paper made, recorded with the paper that made it; it is not a finding of ours, and it is not evidence that the claim is true.

What we can traverse

Step two depends on having somewhere to look. These are the kinds of relationship the assembled substrate can follow between a gene, a disease, a drug and a pathway — the raw material for finding that a mechanism in one condition appears in another.

Counts by type, from the assembled relation substrate. This says what kinds of link exist and how many — never which. The relationships themselves come from licensed third-party sources, are used only in the backend, and are not published here.

What survives each step

And step one is where most of the work goes, because most of it does not survive. A measurement is only usable once it can be compared to another measurement and a model is asking the question. This is the attrition, on one denominator.

Each bar is drawn against the same denominator — every measurement we hold. Extraction is not use: a number is only usable once it can be compared to another number and something is asking the question. Most of what we extract is not yet either, and we would rather show that than the top bar alone.

Why CASK first

Start where it is hardest

CASK-related disorder affects a few hundred known families worldwide. That is not a limitation of the test case — it is the point of it.

The hardest case is the honest one

A method that only works where data is plentiful has not solved the rare-disease problem; it has avoided it. If this works on a few hundred patients and a few dozen papers, it will work wherever the literature is thicker.

It is a test case, not the scope

The substrate is gene-general by construction and already spans thousands of disease genes. Every instrument takes the disease as a parameter. A tool that only worked for CASK would have failed at its actual job.

We are close to it

This programme was built by people with a direct stake in CASK, working alongside researchers and clinicians who study it. That is a reason for the choice, and also a reason for the discipline — wanting an answer is exactly the condition under which you start believing weak ones.

What this is not

The part that has to be said plainly

An ambition stated without its limits is a promise, and we are not in a position to make promises to anyone.

Nothing we produce has been tested in a person. This is in-silico research. Every candidate is a hypothesis awaiting laboratory work, and it becomes meaningful for a patient only after bench and clinical scientists have tested it.
A candidate is not a treatment, and we will not describe one as though it were. Nothing on this site should be acted on medically, and no one should change a child’s care because of anything we have generated.
The generative layers have no proven forecasting skill. They propose; they do not predict. That is precisely why every hypothesis ships with the experiment that would kill it.
We are earlier than the ambition on this page. The instrument that asks what other genes license about CASK currently licenses no empirical transfer onto it — the comparisons that would justify one have not been run in the literature. We would rather publish that than a number that flatters us.
We are not a substitute for a clinician, a genetic counsellor, or a patient organisation. On the things families most need, those people are better than us, and our resources pages point at them rather than competing with them.

Accountability

How you would know if we were wrong

A programme that cannot say what would refute it is not doing science, and this one is built so that the question has answers.

Every prediction carries its falsifier. Each hypothesis names the measurement that would kill it. When a bench result comes back the other way, that is recorded against the prediction rather than quietly forgotten.

The framework is tested against a case where the regulator and the evidence disagreed. A standing check compares two well-known trials of the same disease — the larger, more rigorous one that failed, and the smaller, thinner one that was approved — and requires the framework to score the rigorous trial higher on methodological quality. If regulatory approval ever starts driving our credibility judgments, that check fails and the work stops until it is fixed.

The framework is not allowed to learn from itself. Its own outputs are barred from feeding its own corrective changes — enforced at the code boundary, not by good intentions — so it cannot manufacture agreement with itself and mistake that for evidence.

How the method works What the programme has produced Request access

Research aid only — not medical advice. Piper is a research and educational aid. It is not medical advice, not a diagnosis, and not a clinical determination. Its output is hypothesis-generating and may be incomplete, provisional, or wrong, including AI-generated errors. Always consult your own qualified healthcare professional before making any medical or treatment decision.

The material on this site is written with the help of AI and may contain mistakes. It describes the state of published knowledge rather than any individual, and every source it rests on is named so you can check it.