Automating the Search for Rare Disease Cures

Abstract

More than 7,000 rare diseases have been identified. Individually, each affects a tiny population — often fewer than a few thousand patients worldwide. In aggregate, they represent one of the largest unmet medical needs on the planet. The barriers are structural: small patient cohorts make clinical trials nearly impossible to power, fragmented data makes mechanism discovery slow, and the economics of drug development make most rare diseases commercially unattractive.

Claude Cures runs a continuous, unattended attempt at generating mechanistic hypotheses for these diseases. A foundation model is given the genetic basis, the documented phenotype, the known (or unknown) mechanism, and the therapeutic landscape — and is asked for a checkable increment: a candidate pathway, a druggable target, or a precise characterisation of why current approaches fail. It is explicitly instructed not to claim a cure, and to label every step as [KNOWN], [INFERRED], or [HYPOTHETICAL].

The honest expected outcome is that most hypotheses will not survive expert review. We consider that worth instrumenting anyway. The interesting measurement is not whether the model discovers a treatment — it is whether its reasoning carries signal: whether it identifies mechanistic links between diseases that share a gene or pathway, whether it correctly ranks therapeutic modalities by feasibility, and whether a treated control disease is reconstructed or merely recited.

The discovery protocol

A run does not attempt the disease. It attempts one registered entry from that disease's therapeutic landscape — a specific modality (antisense oligonucleotide, gene replacement, small molecule chaperone, substrate reduction) — and is scored on whether the hypothesis is testable, not on whether it is novel. The instruction that does the most work is the one forbidding a claimed discovery: a model permitted to conclude triumphantly will do so, and the output becomes unfalsifiable speculation. Forced to name the assumption it cannot justify, it produces something a domain expert can evaluate.

For Niemann-Pick Type C, the current entries include cholesterol trafficking restoration via NPC1 stabilisation, substrate reduction targeting sphingolipid synthesis, and pharmacological chaperone approaches. Each run commits to one modality and states why it believes that modality is more tractable than the alternatives.

Every run carries the same failure mode: the model reproduces a known mechanism, hits the known wall, and describes the wall. That is the expected output and it is recorded as such. The rare interesting case is a run that reaches the wall by a route the literature does not document.

The validation control

Spinal Muscular Atrophy (SMA) is in the disease set despite having multiple approved treatments (nusinersen, onasemnogene, risdiplam). It is the control. The path from SMN1 deletion to motor neuron degeneration to SMN2 splicing correction is thoroughly represented in any plausible training corpus, so a model asked to "discover" a treatment can succeed by recall alone — which makes it the one disease where we can distinguish genuine mechanistic reasoning from recitation.

A run is scored as reconstruction only if it derives the exon-skipping mechanism from first principles rather than quoting it, explains why intrathecal delivery was necessary for the first generation, and correctly identifies which therapeutic approaches fail in adult-onset forms. Recitation is common. Reconstruction is not.

Watch it reason in real time

The solver cycles the disease set on its own, continuously. Read the reasoning trace as it arrives. Nothing is filtered and nothing is retried — what the model produces on the first pass is what gets published.