README.md

---
title: RareDx — Rare Disease Diagnostic AI
emoji: 🧬
colorFrom: blue
colorTo: indigo
sdk: docker
app_port: 8501
pinned: false
short_description: Multi-agent AI for rare disease diagnosis
---

# RareDx — Rare Disease Diagnostic AI

A multi-agent clinical AI system that generates differential diagnoses for rare diseases from plain-text clinical notes.

## How It Works

1. **Symptom Parser** — maps free-text symptoms to HPO term IDs using BioLORD-2023 semantic similarity
2. **Graph Search** — traverses the Orphanet/HPO knowledge graph (11,456 diseases, 115,839 phenotype associations)
3. **Vector Search** — BioLORD semantic search over HPO-enriched disease embeddings
4. **RRF Fusion** — merges both rankings via Reciprocal Rank Fusion
5. **Hallucination Guard** — FusionNode filters candidates lacking phenotype evidence

## Example

Paste a clinical note like:
> *"18 year old male, extremely tall, displaced lens in left eye, heart murmur, flexible joints, scoliosis"*

The system returns ranked differential diagnoses with evidence scores, matched HPO terms, and an interactive evidence map.

## Architecture

| Component | Technology |
|-----------|-----------|
| Knowledge graph | Orphanet + HPO (NetworkX JSON) |
| Embeddings | FremyCompany/BioLORD-2023 (768-dim) |
| Vector store | ChromaDB (embedded) |
| API | FastAPI on port 8080 (internal) |
| Dashboard | Streamlit on port 8501 |

## Data Sources

- [Orphanet](https://www.orphadata.com/) — rare disease names, definitions, HPO phenotype associations
- [Human Phenotype Ontology](https://hpo.jax.org/) — 8,701 standardised phenotype terms
- [BioLORD-2023](https://huggingface.co/FremyCompany/BioLORD-2023) — biomedical sentence encoder

## Startup Note

The pipeline loads the BioLORD model and 11,456-disease knowledge graph on first request. Allow ~30 seconds after the Space starts before submitting a note.