Skip to content

utkukose/STING-DSS

Folders and files

NameName
Last commit message
Last commit date

Latest commit

 

History

3 Commits
 
 

Repository files navigation



STING DSS

Drug Repositioning and Synthetic Patient Treatment Decision Support System

for Childhood Acute Lymphoblastic Leukemia

My role: Principal Investigator · System architect · Lead developer

Project Repository · Live System · Publications & Data · My GitHub


About this repository

This repository provides an overview of the STING DSS research project. The full source code, deployment configuration, and technical documentation are maintained at the official project repository:

github.com/tubitaksting

For publications, datasets, and citable outputs, visit the project website:

sting.sdu.edu.tr


What is STING DSS?

STING DSS is a full-stack clinical research platform I designed and built as part of TÜBİTAK 1001 Project No. 123E383 (2023–2026) at Süleyman Demirel University. It integrates five AI/ML methodologies into a single, sequential pipeline for pediatric ALL research — from raw drug–protein interaction data all the way to population-level synthetic patient cohort analysis.

The system is deployed as a production-ready, bilingual (TR/EN) web application accessible to the research team and collaborators worldwide.


My contributions

Research leadership

  • Conceived and directed the overall research programme
  • Designed the five-layer methodological architecture connecting drug repositioning, ODE simulation, genetic algorithm optimisation, GNN digital twin prediction, and GAN synthetic patient generation
  • Developed the unified 5-class risk stratification ontology harmonising NCI, COG, BFM, and SJCRH criteria

System development

  • Architected and implemented the full-stack platform: FastAPI backend, React/Vite frontend, Docker Compose infrastructure
  • Designed and implemented TenDrugALLModel — a 48-dimensional ODE system modelling 10 drugs across 4 treatment phases with RK45 adaptive solver
  • Implemented GNN v2 (GCNConv×2, h=256) digital twin achieving median R² = 0.991 across 8 treatment outcome targets
  • Designed the clean-schema CTGAN architecture eliminating label leakage (GAN-label concordance 0.41%), with post-hoc clinical enrichment pipeline
  • Implemented all four XAI methods: SHAP KernelExplainer, Permutation Importance, Counterfactual search, and GEMEX

Open-source library

  • Developed and published GEMEX v1.2.2 — a Riemannian manifold-based explainability library (pip install gemex) used within the STING pipeline for geodesic sensitivity field analysis

Production deployment

  • Deployed the system to a Natro VPS with Nginx reverse proxy, JWT authentication, bcrypt password hashing, slowapi rate limiting, and hardened API configuration
  • Built bilingual Admin Panel with user management, activity logging, and TÜBİTAK research survey integration

Pipeline overview

Tab 1  Bi-LSTM Drug Repositioning      314,531 ligand–protein pairs (DrugBank · ChEMBL · PubChem · KIBA)
  │                                    Copanlisib: −207.11 kcal/mol · Novobiocin: HSP90 inhibitor
  ▼
Tab 2  Patient & Treatment Setup       10 drugs · 4 phases · BSA · TPMT · Vitamin D · lifestyle factors
  │
  ▼
Tab 3  ODE PK/PD Simulation           TenDrugALLModel · 48-dim state · RK45 · 250 days
  │                                    Sensitivity: Vitamin D 48.1% impact on WBC minimum
  ▼
Tab 4  GA Dose Optimisation           BRR d8: 99.64% · EOI MRD: 3.3×10⁻⁵ · DNR: 149.2 mg/m²
  │
  ▼
Tab 5  GNN Digital Twin + XAI         Median R²: 0.991 · SHAP · Permutation · GEMEX · Counterfactual
  │
  ▼
Tab 6  Synthetic Cohort (GAN)         200 patients · 5-class risk · COP+NOV: Lt→0.000 · WBC +17.5%

Selected technical highlights

Component Detail
ODE engine TenDrugALLModel — 48-dim state vector, 10 drugs, 4 phases, RK45 solver
GA optimisation Population 80 · Generations 200 · 4 simultaneous clinical objectives
GNN architecture GCNConv×2 (h=256, dropout=0.2) · heterogeneous patient graph · k=3 lag
CTGAN design Clean-schema (30 static features) · post-hoc MRD/risk/PI/prognosis pipeline
XAI — GEMEX Geodesic Sensitivity Field · Riemannian manifold · κ = 0.5612 (example patient)
Privacy validation MIA AUC = 0.530 (≈ random) · clinical violation rate = 0%
Deployment Docker Compose · Nginx · Redis · Celery · JWT · bcrypt · slowapi
Languages Bilingual TR/EN · light/dark theme · session management

Tech stack

Backend          FastAPI · Python 3.10+ · PyTorch · SciPy (RK45) · SDV/CTGAN · SHAP · GEMEX v1.2.2
Frontend         React 18 · Vite · Recharts · D3.js · Tailwind CSS
Infrastructure   Docker Compose · Nginx · Redis · Celery · JWT · bcrypt · slowapi

Project team

Name Role
Prof. Dr. Utku Köse (me) · utkukose.com · ORCID Principal Investigator / Developer — SDU / University of North Dakota / VelTech / Universidad Panamericana
Prof. Dr. Gözde Özkan Tükel Researcher / Developer — Süleyman Demirel University
Assist. Prof. Dr. İlhan Uysal Researcher / Developer — Burdur Mehmet Akif Ersoy University
Lect. Osman Ceylan Researcher / Developer — Isparta Applied Sciences University
Lect. Emine Betül Sürücü Researcher / Developer — Süleyman Demirel University

Funding

Supported by the Scientific and Technological Research Council of Türkiye (TÜBİTAK), 1001 Programme, Project No: 123E383 (2023–2026), Süleyman Demirel University.


Links

Full source code & docs github.com/tubitaksting
Live system 37.148.208.183:8091
Publications, datasets & citations sting.sdu.edu.tr
GEMEX library (PyPI) pypi.org/project/gemex
My GitHub profile github.com/utkukose
My website utkukose.com
ORCID 0000-0002-9652-6415

Disclaimer

STING DSS is a research prototype. Outputs have not been validated against real clinical data and must not be used for clinical decision-making. Apache 2.0 License.


This repository is maintained by Prof. Dr. Utku Köse as part of the STING project. For the full codebase, visit github.com/tubitaksting.

Releases

No releases published

Packages

 
 
 

Contributors