EMMA-STRAT Results Browser

Interactive model performance explorer for endometrial cancer molecular subtyping

EMMA-STRAT (Endometrial cancer Multi-omics Machine learning Architecture for STRATification) is a supervised multi-omics machine learning framework that integrates mRNA expression, miRNA expression, and DNA methylation data to classify UCEC (Uterine Corpus Endometrial Carcinoma) molecular subtypes and MSI status. The framework was trained on TCGA (N=433) and validated on two independent CPTAC cohorts (Set-2 N=95, Set-3 N=108). Six classifiers were benchmarked — LightGBM, MLP, Random Forest, SVM, KNN, and GNN — with top-performing models achieving a balanced accuracy of 0.981 (LightGBM, MSI classification, internal) and 0.891 (MLP, genomic subtyping, internal), with strong generalisation to both external validation sets.

6
Classifiers Benchmarked
3
Omics Layers Integrated
2
Independent Validation Cohorts
1,152
Models Optimised & Trained
Browse Results
Model Results

Explore balanced accuracy, F1, and AUC across all model and feature selection combinations for both tasks.

Open
Selected Features

View the top features selected for the MSI and genomic subtype models with their omics layer annotations.

Open
SHAP Analysis

Inspect per-class feature importance scores showing which molecular features drive each subtype prediction.

Open
Flexynesis Comparison

Compare EMMA-STRAT performance against three Flexynesis integration strategies across all evaluation sets.

Open
Methods Summary
Data
  • Training: TCGA UCEC (N=433)
  • External validation: CPTAC Set-2 (N=95), Set-3 (N=108)
  • Omics: mRNA expression, miRNA expression, DNA methylation
  • Targets: TCGA molecular subtypes & MSI status
Models
  • 6 classifiers: LightGBM, MLP, Random Forest, SVM, KNN, GNN
  • Hyperparameter optimisation via Optuna
  • 5-fold stratified cross-validation
  • Feature selection: ANOVA, LASSO, RF, SVM
Validation
  • Bootstrapped 95% confidence intervals
  • Calibration curves & ECE
  • Decision curve analysis
  • Flexynesis multi-omics integration benchmarking