Measured result · not a demo

It built a rare disease diagnostic system that beats human state of the art

Nobody gave it data. Nobody gave it tools. Nobody wrote a line of code. 12 hours later, the AI system it delivered beat a system published in the main Nature journal on public benchmarks.

0123456789
0123456789
h

From zero to delivery

0

Data given to it

0

Human intervention

What it set out to challenge

A system a human team built over an entire research cycle

Rare disease diagnosis is one of the hardest problems in medicine: more than 7,000 rare diseases are known worldwide, each one vanishingly rare, and patients typically spend years bouncing between hospitals before a diagnosis is confirmed. In 2026, a human research team published their multi-agent AI diagnostic system in the main Nature journal.

Human state of the art · human SOTA

DeepRare

A multi-agent rare disease diagnostic system published in Nature, Vol 651 (2026).

Architecture designed by a team of human experts

Went through a full research and publication cycle

Beat every prior method on the benchmarks

Outperformed senior clinicians on some datasets

Our system · built autonomously by AI

Apodex Discovery

One sentence of task description. Zero data. Zero human intervention. All of it done by the AI itself.

Found its own data online, cleaned it, standardized it

Built its own knowledge base and retrieval engine

Wrote its own tools, designed its own experiments

Trained its own models, tuned them, and signed off on them

How it pulled this off

In 12 hours, five things happened at once

Every act below is a process that actually took place during this run.

01

It went looking for the cases itself

It started from zero records. It searched public medical databases online, found confirmed-case repositories published by global rare disease research consortia, extracted each patient’s symptom codes, causal genes and final diagnosis one by one, then filtered out duplicates and low-quality samples.

Monarch Initiativephenopacket-storeConfirmed case reports
Scanned10,606
Ingested5,203
Diseases covered804
Symptom nodes19,404
Relationship edges48,511
Knowledge sources4

02

It built its own medical knowledge base

Cases alone weren’t enough. It downloaded the entire international-standard Human Phenotype Ontology — a vast network describing the hierarchical relationships between every clinical symptom — then wired in disease-to-symptom tables and gene-to-disease mappings to assemble a knowledge foundation it could reason over.

03

It wrote its own tools, then kept rewriting them

Knowledge sitting on a shelf isn’t usable — you need decent instruments to read it. For instance, it built itself a set of “diagnostic instruments”: starting from symptoms, it follows three lines of evidence in parallel to find candidate diseases, scores every candidate on how well it fits, then trains a ranker to fuse those scores into a confidence ranking. It built more than one toolset like this, and every one of them was torn down and rewritten many times over — each version had to be measured to see whether it genuinely improved, and if it didn’t, it was rolled back.

Three candidate channels27-dim clinical featuresConfidence ranker
Versionv4
Feature dims27
Accuracy0.834
Turn3
Tool calls8
Pipeline stateIdle

04

It assembled all of it into a “workbench”

Data, knowledge base and tools were assembled into a complete runtime environment — a workbench where an AI doctor can check and re-check its own work: retrieve the evidence first, verify it against authoritative descriptions, and only then commit to a conclusion. It then ran that pipeline end to end, validated it, and optimized it.

RetrieveVerifyFinalizeNever a blank answer

05

It trained the models, tuned the knobs, and picked the winner

It built and optimized its own training set, launched several runs with different hyperparameters in parallel, and then picked the best performer on a held-out set it had reserved for itself. At no point did anyone tell it what configuration to use, how long to train, or which one to pick.

Self-built training setParallel hyperparameter searchHeld-out selection
Parallel jobs4
Training samples413
Best modelr16 · lr2e-4

The final deliverable

12 hours later, this is what it handed over

Not a single AI model, but a complete diagnostic system — an AI doctor that knows how to use tools, plus every piece of knowledge and instrumentation it needs to work. The whole thing can be packaged up, taken away, and run and evaluated offline.

AI doctor · model

35B model

Lightweight, deployable on-premise — your patient records never have to leave

What it learned is process

When to look things up, what to look up, and when it is safe to conclude

It does not memorize knowledge

All the medical knowledge lives in the knowledge base

Knowledge base

International-standard symptom ontology

Disease-to-symptom mappings

Gene-disease associations

5,203 confirmed cases

Covering 804 rare diseases

Diagnostic instruments

Multi-path candidate retrieval

Clinical fit scoring

Confidence ranker

Fallback path

No failure anywhere leaves you without an answer

Diagnostic process · can loop back at any step

1 · Read patient symptoms2 · Retrieve candidates3 · Score the evidence4 · Verify against references5 · Revise the ranking6 · Output ranked diagnoses

↺ When evidence is thin or in doubt, it returns to step 2 and searches again

Measured results

On the same benchmarks, it beat human state of the art

Standard benchmarks: 556 real cases it had never seen. The evaluation was re-run by a third party in an isolated environment — the AI could not see the questions and received no feedback. The metric is standard Recall@K — the share of cases where the correct diagnosis appears among the system’s top K candidates.

DeepRare · human state of the art (Nature, 2026)

System built autonomously by AI

Recall@1

Right on the first guess

0.628

0.678

+7.9%

Recall@3

Correct answer in the top 3

0.704

0.790

+12.2%

Recall@5

Correct answer in the top 5

0.730

0.822

+12.6%

Micro-averaged over 556 held-out cases drawn from LIRICAL, MyGene2 and MME. Human state of the art uses the values publicly reported in the DeepRare paper.

What this means for you

Your data never has to leave your own server room

In this run, all Apodex Discovery received was one sentence of task description. It found the data itself from public sources, built the system itself, and the packaged deliverable runs the entire evaluation offline. In other words — from start to finish, we never touched a single piece of private data.

Build phase

All you give is one sentence

Spell out what problem you want solved and what counts as getting it right. Data collection, knowledge base, tools and training are all handled by the system itself. Your data plays no part in training.

Deliverable

The whole system is handed to you

A lightweight model plus every piece of knowledge and instrumentation it needs to work, packaged up and deployed on your own servers. Not an API call — an asset you own.

In use

Works with the network unplugged

Inference needs no internet connection at any point. Patient records, samples, molecular structures, experiment logs — this data moves only inside your own server room, and never a step beyond it.

The public world

Public literature and databases

International standard ontologies

Published cases and experiments

The system learns and builds here

Only the system crosses this line

YOUR DATA BOUNDARY

Your data does not cross

Your server room

The delivered purpose-built system

Your private data

Offline inference, no network

Data only circulates in here ↺

One quiet aside: if you are willing to provide your own private data, the system can do even better with it — but that is a bonus, not a requirement.

What this actually means

This was never about winning at rare disease diagnosis

What was actually proven is this: Apodex Discovery can already run an entire research pipeline on its own — setting the approach, finding the data, building the tools, designing the experiments, training the models, validating itself, and delivering the result. Apodex Discovery is now being applied to biology, clinical medicine, chemistry and materials. And all you ever have to give it is one sentence.

0123456789
0123456789
hours

Zero to delivery

0 / 0

Data given · human intervention

0123456789
0123456789
.0
% R@1

vs. human SOTA 62.8%

All figures come from the complete trace log of a single real run and third-party evaluation results. Human state of the art references the values publicly reported in DeepRare (Nature, Vol 651, 2026).