Ilsun Yun Ph.D. Candidate

Role

Led machine learning and clonality analysis

Computational Contributions

• Developed ensemble ML classifier (13 base models: logistic regression, SVM, random forest, XGBoost) to distinguish somatic vs. germline mutations in tumor-only samples, achieving F1=0.89 and enabling analysis of 9 patients without matched normals
• Implemented automated feature engineering pipeline extracting 50+ genomic features (allele frequency, read depth, population databases, functional predictions)
• Performed tumor clonality analysis using FastClone: quantitatively demonstrated CPC's higher intratumoral heterogeneity (62.5% had ≥2 clones vs. 0% CPP, p=0.019)

Impact

First ML-driven quantitative demonstration of clonal complexity differences between benign and malignant choroid plexus tumors; enabled genetic analysis without matched normal samples

Role

Led entire computational and ML analysis

ML Pipeline Development

• Built random forest classifiers for CMS4 prediction achieving AUC=0.931 (TCGA) and Macro-AUC=0.857 (CRCSC multi-platform validation)
• Performed systematic hyperparameter optimization (100+ configurations) reducing overfitting from 41.2% to 13.9% through stratified cross-validation and feature selection
• Developed consensus prediction framework combining multiple models to improve robustness across platforms (RNA-seq, microarray, qPCR)

Model Interpretation & Validation

• Applied SHAP analysis revealing PDGFRB as primary predictive driver (2.36× more important than next feature)
• Validated computational predictions on independent qPCR cohort: consensus approach showed significantly higher PDGFRA/B expression in predicted CMS4 (p<0.001)
• Demonstrated cross-platform generalization: model trained on RNA-seq successfully predicted qPCR-based subgroups

Impact

Created cost-effective qPCR-based CMS4 classifier as alternative to expensive RNA-seq (>$500 vs $50/sample); demonstrated platform-agnostic ML approach applicable to resource-limited clinical settings

Role

Led entire bioinformatics and ML analysis

Pipeline Development

• Built complete automated 16S analysis pipeline (QIIME2): implemented quality control, taxonomic assignment (SILVA database), alpha/beta diversity analysis, and batch effect correction
• Processed >14 million sequences with reproducible workflow achieving 99.2% read retention after quality filtering

Unsupervised ML & Survival Analysis

• Performed unsupervised clustering (hierarchical clustering on Bray-Curtis distances) identifying 3 prognostic microbiome groups with significantly different survival outcomes (Cluster 1 vs. 3: HR=3.2, p=0.00009)
• Demonstrated bile's superiority over stool for prognosis: bile-based clusters discriminated survival (p<0.001) while stool-based clustering showed no association (p=0.42)
• Applied PICRUSt2 functional prediction revealing gingipain protease pathway enrichment in worst-outcome cluster (LEfSe LDA>3.5)

Statistical Validation

• Multivariate Cox regression confirmed microbiome clustering as independent prognostic factor after adjusting for stage, age, treatment (HR=2.8, p=0.003)
• Implemented permutation testing (10,000 iterations) validating cluster stability and ruling out chance findings

Impact

First study establishing bile microbiome community structure as independent prognostic factor; identified potential therapeutic targets (gingipain inhibitors) for high-risk patients

Role

Led entire bioinformatics and survival analysis

Fusion Detection Pipeline

• Developed automated fusion calling pipeline (STAR-Fusion + custom filtering): processed >2TB RNA-seq data with stringent quality filters (≥5 spanning reads, ≥2 split reads, validated splice sites)
• Identified ethnicity-specific SChLAP1:UBE2E3 fusion in 12.5% localized and 15.5% advanced PCa (significantly higher than TCGA Caucasian cohort: 3.2%, p=0.001)
• Validated fusion-driven SChLAP1 overexpression: fusion-positive cases showed 8.4-fold higher SChLAP1 expression (p<0.001)

Clinical Prediction Modeling

• Performed multivariate Cox regression survival analysis demonstrating fusion as independent predictor of biochemical recurrence in localized disease (HR=2.86, 95% CI: 1.18-6.94, p=0.02) and overall survival in advanced disease (HR=2.73, 95% CI: 1.01-7.38, p=0.048)
• Built combined risk model (fusion + PSA) achieving AUC=0.85 for BCR prediction, comparable to established genomic classifiers (Decipher, Oncotype DX)
• Implemented time-dependent ROC analysis validating consistent predictive performance across follow-up periods

Impact

Established clinical utility of ethnicity-specific fusion for risk stratification; demonstrated potential for precision medicine approach in Korean PCa patients; fusion detection method now adopted for clinical testing pipeline