OncoTwin: A Multi-Modal, Treatment-Based Digital Twin for Cancer Recurrence Risk Prediction and Biomarker Discovery

Overview

Breast cancer recurrence rates reach up to 50%, yet existing clinical tools rely on single modality data or repeated medical check-ups. The proposed model, OncoTwin, is a novel recurrence-risk prediction digital twin framework that incorporates four modalities (proteomic, genomic, clinical, and histopathology images) using TCGA-BRCA patients. Its discovery-first and treatment-based architecture overcomes the limitations of previous models that use predefined gene-signatures or focus on single treatment. In the first molecular module, ~400 RPPA proteins were screened using an Elastic CoxNet Regression model. After applying a LASSO-Cox discordance filter, three strong proteins were identified with 5-fold CV C-index of 0.6571: PRAS40_pT246 (HR=2.653), P27/CDKN1B (HR=0.684), and DJ1/PARK7 (HR=0.722). For the second module, 20,155 RNA-sequence genes were filtered to the top 200 using a Log-Rank Pre-filter. A LASSO-Regularized Cox PH model, along with Random Survival Forest ranking, highlighted 18 strong genes (CV C-index: 0.6781), including C15orf52 (HR=1.157), IL12B (HR=0.880), and PCDHA12 (HR=1.054). GRAD-CAM features were extracted from histopathology tiles using ResNet and Multiple Instance Learning with five-fold CV index of 0.887, removing the need for clinically-annotated images. By combining 80 multi-modal features into a Random Survival Forest, OncoTwin achieves a cross-validated c-index of 0.851. The OncoTwin Interface utilizes PyQt5 to create a patient-specific dashboard with real-time Kaplan-Meier Survival Curves. OncoTwin’s ability to accurately simulate recurrence risk longitudinally overcomes previous approaches’ static prediction, setting the high standard for adaptive disease modeling and patient-specific biomarker discovery.

Competition history

CSEF 2026 Medicine & Physiology (Senior Division) · Entry S-15-19

Related projects

ISEF · 2026