Spatiotemporal Oncology Intelligence Engine
OMNIGEN models cancer treatment resistance as a coupled spatiotemporal problem — clonal evolution over time, drug delivery failure over space — and adversarially stress-tests every proposed treatment against the tumor's known biological escape routes before recommending it.
Anchor cohort: LUAD / EGFR-TKI resistance (Phases 0–5).
Research only. OMNIGEN is not a medical device and must not be used to inform clinical decisions. See Research Disclaimer.
╔══════════════════════════════════════════════════════════════════════╗
║ LAYER 0 — CLONAL DYNAMICS ENGINE ║
║ ║
║ Longitudinal VAF time series (TRACERx / GENIE BPC) ║
║ Stochastic Lotka-Volterra clonal competition model ║
║ Bayesian inference (pymc NUTS) → per-clone growth rate posteriors ║
║ Trigger: resistant clone selection coefficient P95 > threshold ║
╚═══════════════════════════╦══════════════════════════════════════════╝
│ TRIGGER: clone under positive selection
▼
╔══════════════════════════════════════════════════════════════════════╗
║ LAYER 1 — TUMOR MICROENVIRONMENT GRAPH ║
║ ║
║ Spatial transcriptomics: 10x Visium HD / Xenium Prime 5K ║
║ Cell type deconvolution: cell2location (scvi-tools) ║
║ Spatial domain detection: cellcharter ║
║ Heterogeneous PyG graph: tumor · immune · vascular nodes ║
║ ║
║ ┌─────────────────────────────────────────────────┐ ║
║ │ XPINN Physics Solver (deepxde) │ ║
║ │ ∂C/∂t = ∇·(D(x)∇C) − λ(x)C − R(x,C) │ ║
║ │ Drug concentration field — domain decomposition │ ║
║ └─────────────────────────────────────────────────┘ ║
╚═══════════════════════════╦══════════════════════════════════════════╝
│ spatial drug concentration field
▼
╔══════════════════════════════════════════════════════════════════════╗
║ LAYER 2 — ADVERSARIAL MULTI-AGENT CORE ║
║ ║
║ Blue Team: ║
║ Clonal Pathologist — clone dynamics analysis (Haiku) ║
║ Spatial Pharmacologist — diffusion failure analysis (Haiku) ║
║ ║
║ Red Team: ║
║ Tumor Adversary — OmniPath escape route query (Sonnet) ║
║ Biology Constraint Validator — deterministic graph lookup ║
║ ║
║ Synthesis Agent — Pareto protocol generation (Sonnet) ║
╚═══════════════════════════╦══════════════════════════════════════════╝
│ PROPOSED PROTOCOL
▼
╔══════════════════════════════════════════════════════════════════════╗
║ LAYER 3 — TREATMENT VALIDATION ENGINE ║
║ ║
║ Forward simulation: proposed protocol through Layers 0 + 1 ║
║ Gate: resistant clone growth rate P95 must become negative ║
║ If gate fails → protocol rejected back to Blue Team ║
╚══════════════════════════════════════════════════════════════════════╝
# 1. Copy environment config
cp .env.example .env
# Edit .env — set POSTGRES_PASSWORD, REDIS_PASSWORD, ANTHROPIC_API_KEY
# 2. Start services and install dependencies
make dev
# 3. Apply database schema
make migrate
# 4. Verify the integrity suite passes
make test-integrity
# 5. Run the full test suite
make testServices: TimescaleDB on localhost:5433, Redis on localhost:6380, Prefect UI at http://localhost:4201.
| Source | Layer | What It Provides | Access |
|---|---|---|---|
| 10x Genomics public datasets | 1 | Visium HD + Xenium LUAD spatial transcriptomics | Free — no DUA |
| TRACERx (Francis Crick / EGA) | 0 | Multi-region serial liquid biopsy, LUAD longitudinal VAF | DUA required |
| GENIE BPC NSCLC v2.0 (Synapse) | 0, Val | 1,846 NSCLC patients, pre/post-treatment paired samples | DUA required |
| TCGA-LUAD (GDC Portal) | 0, Val | Bulk RNA-seq, WES mutation calls, CNV, clinical data | Controlled access |
| OmniPath (local PostgreSQL) | 2 | Integrated signaling network: Reactome + SIGNOR + STRING | Free |
| TEMULATOR (synthetic) | 0 | Simulated clonal evolution for bootstrap (no DUA needed) | Free |
| Anthropic API | 2 | Multi-agent debate (Haiku + Sonnet) | API key required |
Submit TRACERx EGA access application on Day 1 of Phase 0. Applications take 4–8 weeks.
| Phase | Duration | Description | Gate Criterion | Status |
|---|---|---|---|---|
| 0 | 2 weeks | Environment + data foundations | assert_clinical_integrity 100% coverage; OmniPath ≥ 500 nodes; TEMULATOR loaded |
✅ Complete |
| 1 | 4 weeks | Clonal Dynamics Engine | LV identifies resistant clone in 7/10 blinded TEMULATOR holdout patients | ✅ Complete (model validated via direct diagnostic; gate script limitation documented as D-43) |
| 2 | 5 weeks | TME Graph + Physics Solver | XPINN L2 error < 15% vs FEniCSx; TME graph all 3 edge types | 🔄 In Progress |
| 3 | 5 weeks | Adversarial Multi-Agent Core | Debate completes 3/5; BCV rejects 16/20 fabricated claims | Pending |
| 4 | 4 weeks | Integration + Forward Simulation | Direction correct 15/20; no temporal leakage | Pending |
| 5 | 4 weeks | Clinical Validation | Onset MAE ≥ 20% below FLAURA baseline; mechanism accuracy ≥ 50% | Pending |
If you use OMNIGEN in research, please cite:
@software{omnigen2026,
author = {shiviancode},
title = {{OMNIGEN}: Spatiotemporal Oncology Intelligence Engine},
year = {2026},
url = {https://github.com/shiviancodes/omnigen},
license = {Apache-2.0},
}Key dependencies to cite: pymc (Salvatier et al. 2016), deepxde (Lu et al. SIAM Review 2021), cell2location (Kleshchevnikov et al. Nature Biotechnology 2022), PyTorch Geometric (Fey & Lenssen 2019), pymoo (Blank & Deb 2020), OmniPath (Türei et al. 2021).
OMNIGEN is a research prototype intended for computational oncology research only.
- It is not a medical device, clinical decision support system, or diagnostic tool.
- It is not validated for clinical use and must not influence patient care decisions.
- All results are based on computational models and retrospective data analysis.
- Any clinical interpretation of outputs requires independent validation by qualified oncologists with access to complete patient records.
- Use of TRACERx, GENIE BPC, and TCGA-LUAD data is subject to the respective data use agreements. Users are responsible for compliance.
This software is provided "as is" under the Apache 2.0 License, without warranty of any kind. See LICENSE for details.