Author
Adhaar Kohli
Degree
Master of Technology (M.Tech) • 2026
Institution
Jaypee Institute of Information Technology, Noida, India
This thesis develops LungVar-ML, an integrated bioinformatics pipeline that translates raw multi-omic data into a translatable biosensor for non-small cell lung cancer (NSCLC) biomarker detection. The pipeline harmonizes transcriptomic, somatic mutation, methylation, and copy-number data from 982 TCGA-NSCLC tumors. A consensus procedure reduces approximately 50,000 features to a panel of ~250 biomarkers. A two-level stacking ensemble achieves a cross-validated AUC of 0.913 and a concordance index of 0.821, outperforming other algorithms and retaining discrimination on external cohorts. Interpretability analysis identifies a proliferation-immune axis and links to STK11/KEAP1 co-mutation. The panel is translated into a circulating-miRNA biosensor design with high sensitivity, achieving a diagnostic AUC of 0.958.
Internal AUC-ROC
0.913
C-index
0.821
Biosensor AUC
0.958
Log-rank p-value
<0.0001
Pan-Omic Data Acquisition
Gathering data from TCGA (LUAD/LUSC), GEO, and cBioPortal — 982 tumors, ~50,000 features across 5 omic layers.
Quality Control & Batch Correction
Standardizing data using ComBat for technical variation correction.
Per-Omic Feature Engineering
Processing RNA-seq, WES/MAF, Methylation, CNV, Clinical data with omic-specific tools.
Consensus Feature Selection
Four independent selectors (mRMR, LASSO, Boruta, Cox-PH) with majority vote → ~250 features.
Two-Level Stacking Ensemble Training
5 base learners (RF, XGBoost, SVM-RBF, ElasticNet, LightGBM) + LogReg meta-learner, stratified k-fold CV, SMOTE.
Risk-Score Generation & Patient Stratification
Calibrated risk scores stratify patients into low/intermediate/high risk groups.
Internal & External Validation
Hold-out sets + two independent GEO cohorts for transportability.
Tumor Microenvironment Deconvolution
CIBERSORTx and ESTIMATE scores for immune profiling.
Interpretation & Biosensor Translation
SHAP interpretability + Si-NWFET biosensor design for miRNA detection.
Transcriptomics (RNA-seq)
Genomics (WES)
Epigenomics (Methylation)
Copy Number Variations
Circulating miRNAs
Stacking Ensemble (Random Forest, XGBoost, SVM-RBF, ElasticNet, LightGBM + LogReg meta)
Consensus Feature Selection (mRMR, LASSO, Boruta, Cox-PH)
Bayesian Hyperparameter Optimization (Optuna)
SMOTE & Cost-sensitive Class Weighting
SHAP & LIME Interpretability
Survival Analysis (Kaplan-Meier, Cox-PH, C-index, Time-dep AUC)
PCA & UMAP Dimensionality Reduction
Molecular Docking (AutoDock Vina)
Molecular Dynamics (GROMACS, MM-PBSA)
Network Analysis (STRING, Cytoscape, MCODE, GO, KEGG, Reactome)