HBVNet

Inspiration

HBV is one of the leading causes of liver cancer worldwide, and distinguishing HBV-related HCC from non-viral HCC is vital for targeted therapies. I wanted to build a clinically available tool that leverages publicly available data (TCGA-LIHC) to show that even with a small, interpretable set of genes, we can achieve strong classification performance.

What it does

HBVNet predicts whether a patient’s HCC is likely HBV-related or non-viral using:

• Clinical features (age, gender, tumor stage/grade)
• RNA-seq expression data
• A minimal 10–20 gene signature selected via LASSO for interpretability

It outputs:

• Predicted probability of HBV-related HCC
• Interactive Gradio app for instant predictions
• Figures (ROC curve, confusion matrix, feature importance)

How we built it

1) Data Acquisition: Downloaded TCGA-LIHC clinical and expression data (FPKM-UQ). 2) Preprocessing:

• Harmonized patient barcodes, merged datasets
• One-hot/ordinal encoded categorical variables
• Imputed missing values 3) Feature Engineering:
• Selected top 500 most variable genes
• Applied L_1-regularized logistic regression to identify minimal gene set 4) Modeling:
• Balanced classes with SMOTE
• Trained and tuned XGBoost models
• Compared clinical-only, expression-only, combined, and minimal signature modalities 5) Evaluation:
• Calculated ROC-AUC, F1-Score, accuracy, plotted ROC curve and confusion matrix 6) Deployment:
• Built an interactive Gradio app with pre-filled mean expression values

Challenges we ran into

• Data harmonization: Matching TCGA barcodes across clinical and expression datasets
• Class imbalance: Required SMOTE and careful evaluation using stratified splits
• Feature dimensionality: > 60,000 genes → had to filter & regularize carefully
• Deployment: Ensuring model artifacts (pkl files) match the app’s expectations

Accomplishments that we're proud of

• Built a reproducible ML pipeline in one week
• Achieved strong performance (accuracy ≈ 0.85 with minimal signature)
• Reduced to a compact gene signature with minimal loss of accuracy
• Deployed a working web app for real-time inference

What we learned

• Feature selection for model interpretability
• Class imbalances in biomedical datasets
• Deploying ML models in a user-friendly interface with Gradio

What's next for HBVNet

• Prospective use: Deploy as a decision support tool for clinicians
• Expand modalities: Integrate mutation data & survival analysis
• Production: Wrap as a REST API or Streamlit dashboard for wider access

Built With

• gradio
• lasso
• machine-learning
• natural-language-processing
• scikit-learn
• smote
• xgb