Raras-AI/gemeo-twin-stack

GEMEO Twin Stack — Application Layer for the GEMEO Patient World Model

The graph-native digital-twin application stack that runs on top of the GEMEO World Model. Six inference modes — trajectory, diagnosis, risk/survival, counterfactual, repurposing, cohort — wired together with auxiliary heads, KG embeddings, and a FastAPI surface. Research preview. Not a medical device.

Authors: Raras.ai team · Contact: dimas@raras.ai Source: github companion repo (raras.org) Paper v1 (Zenodo): DOI 10.5281/zenodo.20092131 🌍 World Model (the dynamics core this stack runs on top of): Raras-AI/gemeo-world-model 📱 Mobile decision-support sibling: Raras-AI/araras-gemma4 License: CC-BY-NC 4.0 + non-clinical-use rider (see LICENSE)

Note on naming. Previously released as Raras-AI/gemeo-world-model (HF auto-redirects). Renamed to gemeo-twin-stack because the actual world model — the generative dynamics core — is the Causal Diffusion Forcing transformer now living at the new gemeo-world-model slug. This repo is the application layer (encoder, cohort, risk, whatif, repurpose, ask, ground_sus, api) that turns the world model into a usable digital twin with six inference modes.

---

What this is

This is the application layer of GEMEO — a module (≈22k LOC) that takes the GEMEO World Model (a Causal Diffusion Forcing transformer, published separately at Raras-AI/gemeo-world-model) and wires it together with auxiliary heads, KG embeddings, and tooling to produce a complete digital-twin product. Each inference mode has a clean Python API, a bootstrap implementation that runs today, and an optional learned slot that drops in when a checkpoint exists.

GEMEO Twin Stack  (this repo)
├── Patient embedding       (gemeo/encoder.py)        HGT scaffolded; bootstrap = weighted KG embedding
├── Cohort retrieval        (gemeo/cohort.py)         kNN + Cypher overlap on PrimeKG
├── Subgraph reasoning      (gemeo/subgraph.py)       1-hop sparsification (learned variant in train/)
├── Trajectory mode         (gemeo/trajectory.py)     ← calls into the GEMEO World Model
├── Diagnostic mode         (deeprare_diagnosis.py +  ← multi-agent over PrimeKG paths
│                            fast_dx.py, ensemble)
├── Risk / survival         (gemeo/risk.py)           NeuralSurv trained on DATASUS SIM (c-index 0.70)
├── Drug repurposing        (gemeo/repurpose.py)      TxGNN slot + SUS auxiliary head
├── Counterfactual          (gemeo/whatif.py)         do-calculus mutator; uses world-model rollout
├── Active learning         (gemeo/ask.py)            info-gain over KG annotations
├── SUS grounding           (gemeo/ground_sus.py)     PCDT/CEAF/UF coverage check
└── FastAPI                 (gemeo/api.py)            production /api/gemeo/* endpoints

GEMEO World Model  (separate repo: Raras-AI/gemeo-world-model)
└── Causal Diffusion Forcing transformer (19.86M params)
    The generative dynamics core. Trajectory + counterfactual modes
    above delegate to it.

---

Headline numbers (real, audited)

Mode	Metric	Value	Notes
Diagnosis (RareBench, 200 cases, v49)	R@1 canonical	57.0% (114/200)	matches DeepRare phenotype-only published level
	R@5 canonical	81.5%
	R@1 strict (ORPHA-code)	39.5%	strict pattern matching
	timeout rate	0.0%
Risk / survival (NeuralSurv, DATASUS SIM)	val c-index	0.70 (best @ ep 30)	4,624 SIM mortality records, 37,494 censoring samples
	final c-index	0.694	after 100 epochs
World Model backbone (gemeo-world-model)	val cross-entropy	0.030	per-token diffusion forcing
	calibration ICI	0.0006	well calibrated
	training	5.8 min on 1×H100	$0.48

These are the audited, end-to-end numbers from the same swarm-py module — they are NOT promises or projections.

---

Six inference modes

1 · Trajectory (digital twin)

from gemeo import build_gemeo

twin = await build_gemeo(
    case_text="5-year-old boy, progressive ataxia, telangiectasia, elevated AFP.",
    patient_info={"age": 5, "sex": "M"},
    context={"sus_region": "SP"},
)
twin.trajectory.horizons   # 6/12/24-month event predictions

Backbone: bootstrap LLM-over-natural-history today, Causal Diffusion Forcing when the CDF checkpoint is mounted.

2 · Diagnosis

twin.diagnoses[:5]   # top hypotheses ranked, with evidence chain

Three diagnostic engines wired in:

• deeprare_diagnosis.py (DeepRare-style multi-agent over PrimeKG paths)
• fast_dx.py (low-latency phenotype → ORPHA ranking)
• ensemble_diagnostic.py (ensemble over the above)

Baseline RareBench numbers reported above. v49 corresponds to the checkpoint shipping in this repo.

3 · Risk / survival

twin.risk.survival_curve   # months → P(alive), with bootstrap CI

artifacts/neuralsurv.pt (186 KB) — Bayesian-style neural survival head over KG-walk features. Trained on DATASUS SIM (n=4,624 deaths, n=37,494 censoring samples). See artifacts/neuralsurv_datasus_summary.json for the per-ORPHA sanity check (e.g., ORPHA:586 / cystic fibrosis: 96.4% alive at 12 m, 80.3% at 72 m).

4 · Counterfactual (what-if)

twin.what_if({"add_treatment": "ivacaftor"})   # mutated trajectory

Bootstrap = heuristic snapshot mutation; learned variant = CF-GNNExplainer slot.

5 · Drug repurposing

twin.drugs.candidates[:5]   # KG-walk Disease→Gene→Drug, SUS-availability ranked

TxGNN slot + SUS auxiliary head (filters by CONITEC/PCDT coverage).

6 · Cohort + active learning

twin.cohort.members[:10]      # patients-like-mine (kNN on PrimeKG)
twin.next_questions[:3]       # information-gain ranked clinical questions

---

What's in this repo

src/gemeo/                            # 22 k LOC, the world-model module
├── core.py, api.py, types.py         # orchestrator + FastAPI + types
├── encoder.py, cohort.py, subgraph.py
├── trajectory.py, risk.py, whatif.py
├── repurpose.py, ask.py, ground_sus.py
├── deeprare_diagnosis-style modules  # diagnostic mode wiring
├── train/                            # HGT, TGNN, TxGNN, NeuralSurv training
├── datasus/                          # DATASUS pull (SIH, APAC, SIM, CNS linkage)
└── cwm/                              # earlier causal-world-model exploration

artifacts/
├── neuralsurv.pt                     # 186 KB, NeuralSurv risk head (c-index 0.70)
├── neuralsurv_datasus_summary.json   # per-ORPHA survival sanity check
└── dt_fm.pt                          # 7.6 MB, decision-time foundation-model baseline

data_derived/                         # ALL derived from PrimeKG (CC-BY 4.0)
├── fused_embeddings_fp16.npz         # 43 MB, 3072-dim disease/HPO/gene embeddings
├── graph_embeddings.npz              # 7 MB, 768-dim KG node embeddings
├── hetero_graph.json                 # 6 MB, edge index by relation
└── node_ids.json                     # 530 KB, id → ORPHA/HPO/HGNC mapping

benchmarks/
└── rarebench_v49_diagnose_200.json   # 200-case RareBench eval, per-case results

examples/
└── quickstart.py                     # minimal usage

LICENSE
README.md   ← this file

The GEMEO World Model (the Causal Diffusion Forcing dynamics core) is published separately at Raras-AI/gemeo-world-model (~80 MB) — keep that repo for the heavy ckpt, this repo for the application stack that runs on top of it.

---

Quickstart

import sys; sys.path.append("src")
from gemeo import build_gemeo

twin = await build_gemeo(
    case_text="Menino, 5 anos, ataxia progressiva, telangiectasia, AFP elevado.",
    patient_info={"age": 5, "sex": "M"},
    context={"sus_region": "SP"},
)

twin.diagnoses[:3]              # → e.g. Ataxia-Telangiectasia (ORPHA:100) at top
twin.cohort.members[:5]         # patients-like-mine
twin.risk.survival_curve        # months → P(alive)
twin.drugs.candidates[:3]       # repurposing
twin.next_questions[:3]         # active learning
twin.sus_check.pcdt_url         # PCDT compliance link

For the FastAPI server (production layout — same as raras.org runs):

pip install -r requirements.txt
uvicorn gemeo.api:app --reload --port 8000
# POST /api/gemeo/build           — create a twin
# POST /api/gemeo/{id}/evolve     — add new clinical data
# POST /api/gemeo/{id}/whatif     — counterfactual
# GET  /api/gemeo/{id}/{cohort,subgraph,trajectory,risk,drugs,trials,next-questions,sus,viz}

---

Data, ethics, governance

• Source: Brazilian DATASUS subsystems (SIH-RD, APAC-SIA, BPA-I, SIM).
• Linkage: CNS-hash deterministic (Tier 1 via APAC).
• De-identification: ages bucketed (5y); UF only (no município); k-anonymity ≥ 5.
• Ethics: Brazilian Res. CNS 466/2012 + 510/2016. LGPD-compliant.
• Not on this repo: PHI, raw CNS hashes, individual-level data. Only derived embeddings (PrimeKG), trained weights (NeuralSurv, DT-FM), source code, and aggregate metrics.

---

What's in development (and what Mayo would enable)

The architecture is mode-complete: every box has a working bootstrap or trained slot. The bottleneck is data substrate:

Mode	SUS today	What Mayo / multimodal substrate unlocks
Trajectory	events only (no notes)	+ 1.65 B clinical notes → richer event tokens
Diagnosis	R@1 57% canonical / 200 cases	+ WES variants + HPO from notes → target R@1 ≥ 78% (DeepRare bar)
Risk	c-index 0.70 on 4,624 deaths	+ longitudinal labs + meds → target c-index ≥ 0.80
Counterfactual	heuristic	+ trial-emulation labels → causal-grade what-if
Repurposing	TxGNN + SUS heuristics	+ Mayo trial outcomes → real evidence ranking

This is the basis of our Mayo Clinic Platform_Accelerate proposal: same world model, multimodal substrate.

---

Citation

@misc{gemeo_world_model_2026,
  title  = {GEMEO: A Patient World Model for Rare Disease, Grounded in
            Brazilian SUS Data},
  author = {Timmers, Dimas and the Raras.ai team},
  year   = {2026},
  url    = {https://huggingface.co/Raras-AI/gemeo-world-model},
  note   = {Research preview. Not a medical device.}
}

@misc{gemeo_v1_2026,
  title  = {GEMEO v1: A SUS-Grounded Patient Digital Twin for Rare-Disease
            Trajectory Forecasting},
  author = {Timmers, Dimas and Kawassaki, Alexandre},
  year   = {2026},
  doi    = {10.5281/zenodo.20092131},
  url    = {https://doi.org/10.5281/zenodo.20092131}
}

Building blocks (please also cite as appropriate)

• PrimeKG — Chandak, Huang, Zitnik. Nature Sci Data (2023).
• HGT — Hu et al., WWW 2020. Heterogeneous Graph Transformer.
• TGNN — Rossi et al., ICML 2020. Temporal Graph Networks.
• TxGNN — Huang et al., Nature Medicine 2024. A foundation model for clinician-centered drug repurposing.
• NeuralSurv — Lee et al., Stat. Med. 2021.
• Diffusion Forcing — Chen et al., NeurIPS 2024 (2407.01392).
• DeepRare — Nature 2026, s41586-025-10097-9.
• PhenoKG — arXiv 2506.13119 (Jun 2025).
• MEDS — Medical Event Data Standard v0.4.1, McDermott et al. 2024.

---

Changelog

• 2026-05 (this release) — Initial public release of the twin-stack application layer + NeuralSurv ckpt (c-index 0.70) + DT-FM baseline + PrimeKG-derived embeddings + RareBench v49 results.
• 2026-05-19 — Renamed from Raras-AI/gemeo-world-model → Raras-AI/gemeo-twin-stack (HF auto-redirects). The slug gemeo-world-model now hosts the actual Causal Diffusion Forcing world model.
• 2026-05 — GEMEO World Model v2 (CDF backbone) published at Raras-AI/gemeo-world-model.
• 2026-04 — GEMEO v1 paper published on Zenodo (DOI 10.5281/zenodo.20092131).

---

⚠️ Reminder: Research only. Not a medical device. No clinical use without physician oversight and applicable regulatory clearance.