OncoLnc-BRCA: AI lncRNA Prognosis Platform

💡 Inspiration

Breast cancer remains the most diagnosed malignancy in women worldwide. While clinical staging and standard protein biomarkers (ER, PR, HER2) guide therapy, patient outcomes remain frustratingly heterogeneous. Two patients with identical clinical stages often experience radically different survival trajectories.

While 98% of the human genome is non-coding, conventional diagnostics largely ignore Long Non-Coding RNAs (lncRNAs) — the hidden "master regulators" of oncogenesis, drug resistance, and immune evasion. Our team (combining medical science and AI) was inspired to bridge this gap: Could we harness machine learning and genomic transcriptomics to uncover prognostic non-coding RNA drivers and build an accurate survival risk prediction platform for breast cancer?

🧬 What It Does

OncoLnc-BRCA is an end-to-end AI-powered precision oncology platform that decodes patient transcriptomes to predict breast cancer survival risk and uncover therapeutic targets.

• Transcriptomic Feature Discovery: Scans genome-wide RNA sequencing data across 1,073 breast cancer patients, filtering out protein-coding noise to pinpoint prognostic lncRNAs.
• Prognostic Risk Stratification: Computes an individualized multi-gene risk score: $$ \text{Risk Score} = \sum_{k=1}^{m} \beta_k \cdot \log_2(\text{TPM}_k + 1) $$ Stratifying patients into High-Risk vs. Low-Risk cohorts with a 3.36-fold increased hazard of mortality ((p = 5.65 \times 10^{-11})).
• Multi-Model Machine Learning: Benchmarks 6 diverse AI classifiers to predict patient survival horizons.
• Biological & Pathway Mapping: Connects identified non-coding biomarkers to validated cancer pathways (NOTCH/NFIB, Wnt/$\beta$-catenin, Hippo/YAP, and PD-1/CTLA-4 immune checkpoints).
• Interactive Clinical Dashboard: Provides a zero-dependency web interface allowing oncologists and researchers to simulate patient survival risk in real time.

🛠️ How We Built It

We structured our bioinformatics and machine learning pipeline into a modular 11-step architecture:

1. Transcriptome Extraction & Annotation: Processed 1,118 raw STAR-Counts files from TCGA-BRCA, isolating 16,901 unique lncRNAs based on GENCODE v36 annotations.
2. Clinical Alignment: Queried the Genomic Data Commons (GDC) API to obtain clinical follow-up and survival endpoints, establishing a 1-to-1 matched cohort of 1,073 primary solid tumors.
3. Quality Control & Variance Stabilization: Filtered low-expression background noise ((\text{TPM} \ge 0.5) in (\ge 10\%) of patients) down to 5,886 actively expressed lncRNAs, transformed with (\log_2(\text{TPM} + 1)).
4. Genome-Wide Survival Modeling: Executed Univariate Cox Proportional Hazards regression and Mantel-Cox Log-Rank tests across all 5,886 genes, discovering 450 statistically significant prognostic lncRNAs ((p < 0.05)).
5. Multi-Factorial Prioritization (CPS): Formulated a Composite Prioritization Score integrating statistical significance, hazard effect sizes ((|\ln(\text{HR})|)), and ML feature weights to prioritize the top 12 candidate biomarkers (e.g., oncogenic LINC01235, STK4-AS1, AC007493.1, and protective MAPT-AS1, LINC01871, ST7-AS1).
6. Machine Learning Pipeline: Trained and validated 6 supervised classifiers (Random Forest, Gradient Boosting, ElasticNet Logistic Regression, SVM, Neural Network MLP, and Soft-Voting Ensemble) using Stratified 80/20 train-test splits and 5-Fold Cross-Validation.
7. Interactive Web Platform: Built both a responsive web dashboard (HTML5/TailwindCSS/JS) and a Python Streamlit application with interactive sliders and real-time clinical risk calculators.

⚡ Challenges We Faced

• Data Imbalance & Endpoint Selection: The raw metastasis clinical data was only 2% complete, making direct metastasis classification statistically invalid. We pivoted to overall survival (OS), which was 99.9% complete across 1,073 patients.
• High-Dimensional Genomic Noise: Filtering out 11,000+ unexpressed transcripts without losing rare but vital regulatory lncRNAs required multi-tiered filtering thresholds.
• Preventing ML Overfitting on Survival Endpoints: Unpenalized tree algorithms struggled with class imbalance (14% mortality vs 86% survival). We engineered balanced class weights and regularized ElasticNet regression, achieving a robust Cross-Validation AUC of 0.690 ± 0.045 and Test ROC-AUC of 0.661.

🏆 Accomplishments We're Proud Of

• Discovered a 12-lncRNA prognostic signature with strong statistical separation ((\text{Hazard Ratio} = 3.36, p = 5.65 \times 10^{-11})).
• Identified key therapeutic targets such as LINC01235 (linked to NOTCH signaling and chemoresistance) and LINC01871 (strongly correlated with cytotoxic CD8+ T-cell infiltration and PD-1/CTLA-4 immune checkpoints).
• Built a fully automated, reproducible 11-step pipeline and a standalone interactive clinical simulation dashboard.

📚 What We Learned

• Non-coding RNAs are potent prognostic engines: lncRNAs carry immense diagnostic signal that rival or complement traditional protein-coding panels.
• The power of cross-disciplinary collaboration: Bridging medical biology with data science and machine learning is crucial for transforming raw genomic data into actionable clinical insights.

🚀 What's Next for OncoLnc-BRCA

• Wet-Lab RT-qPCR Validation: Validating the top 12 lncRNA expression panel on prospective clinical breast cancer biopsy samples.
• Liquid Biopsy Translation: Developing plasma exosomal assays for non-invasive blood-based disease monitoring.
• Pan-Cancer Expansion: Extending our pipeline to identify prognostic lncRNAs in other high-mortality malignancies (e.g., ovarian, lung, and pancreatic cancers).

Built With

• bioinformatics
• breast-cancer
• cancer-research
• cox-regression
• data-science
• elasticnet
• kaplan-meier
• lncrna
• logistic-regression
• machine-learning
• python
• survival
• tcga