About — MEDICAT

§1

The mission

More than 7,000 rare diseases have been identified. In aggregate, they affect an estimated 300 million people worldwide. Yet over 90% have no approved treatment — not because the science is intractable, but because the economics of drug development make most rare diseases commercially unattractive, and the patient populations are too small to power conventional clinical trials.

MEDICAT is a bet that continuously running AI inference — pointed at mechanistic biology and forced to label every claim — can generate enough signal to be useful. Not a cure. Not a shortcut. A checkable increment: a candidate pathway, a druggable target, or a precise characterisation of why current approaches fail.

Every transcript is published unedited. The failures, which constitute nearly all of the output, are as informative as the successes.

§2

How it works

MEDICAT runs an unattended loop. Every few hours, the system selects a disease from the tracked set, pulls enriched profile data from scientific databases, and sends a structured prompt to Claude — Anthropic's frontier AI model — asking for a mechanistic hypothesis focused on one specific therapeutic modality.

The output is constrained: the model is forbidden from claiming a discovery or a cure. It is required to label every statement as [KNOWN], [INFERRED], or [HYPOTHETICAL] — and to explain its reasoning chain so a domain expert can check it step by step.

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.

The completed session is stored in a database and published immediately to the Research Log — no editing, no filtering. You can watch it happen in real time on the Live page.

§3

Label taxonomy

Every claim in a transcript is tagged with one of three labels:

        [KNOWN]        Established in peer-reviewed literature. Verifiable
               against OMIM, ClinVar, or cited publications.

[INFERRED]     Logically derived from known facts but not explicitly
               stated in the literature. The reasoning chain is
               shown and can be checked step by step.

[HYPOTHETICAL] A novel claim generated by the model. Not established
               and not directly derivable from documented evidence.
               These are the potentially interesting outputs — and
               the ones most likely to be wrong.
      

The majority of output in any given run is [KNOWN] — the model restating established biology. This is expected and necessary: the interesting measurement is whether the small fraction of [HYPOTHETICAL] claims carry any signal that a domain expert finds worth investigating.

§4

Data sources

Disease profiles are dynamically enriched using established biomedical databases. Each profile connects to real-time APIs to pull verified identifiers, ontology terms, and cross-species genotype–phenotype links.

Monarch Initiative API (v3) — MONDO ontology IDs, cross-species genotype–phenotype links, and Human Phenotype Ontology (HPO) terms
OMIM — Online Mendelian Inheritance in Man: locus and allele cross-references
GARD — Genetic and Rare Diseases Information Center identifiers
Orphanet — Rare disease nomenclature and expert classification
ClinVar — Genomic variation and clinical significance

We collaborate with the Monarch Initiative and intend to contribute outputs to their public knowledge graph for the broader research community. We are also evaluating integration with Monarch DisMech for automated mechanistic disease classification.

§5

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 the model's training data — which makes it the one disease where we can distinguish genuine mechanistic reasoning from recitation.

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

The control calibrates our confidence in outputs for unsolved diseases. If the model cannot reconstruct the logic for a disease whose answer is known, its hypothetical outputs for unsolved diseases should be weighted accordingly.

§6

Honest limitations

This system has significant limitations that we state plainly:

The model cannot perform experiments. Every hypothesis requires wet-lab validation.
The model's knowledge is limited to its training data. It cannot access recent publications, unpublished results, or proprietary datasets.
For ultra-rare diseases (prevalence <1 in 1,000,000), the training data may contain insufficient information for meaningful hypothesis generation.
The model may confabulate plausible-sounding but incorrect biology. The label taxonomy helps but does not eliminate this risk.
Cross-disease mechanistic insights — the most valuable potential output — require domain expert validation that we are not yet equipped to provide systematically.
This system does not address the diagnostic odyssey, insurance barriers, or access to care infrastructure that affect rare disease patients.

We believe the right response to these limitations is transparency, not silence. Every transcript is published unedited. The failures are as informative as the successes.

§7

Technology

MEDICAT is built on a small, deliberately simple stack. The goal is to keep the inference loop running continuously with minimal operational overhead.

Inference engine: Claude Sonnet 5 (Anthropic) — frontier reasoning model
Backend: Node.js server on Render, persistent service
Database: MongoDB — stores completed research sessions and transcripts
Disease data: Monarch Initiative API, enriched at runtime per session
Funding: $MCAT creator revenue on Robinhood Chain
Source: github.com/sp3aker2020/Medicat

MEDICAT is built with support from Anthropic's AI for Science program — the rare disease research grants track, which provides API credits for researchers working on rare genetic diseases.

About MEDICAT