Cancer Multi-Omics Intelligence Platform

Built for NeoHacks GIKI Edition 2025. An end-to-end, real (not templated) multi-omics ML pipeline + Streamlit app trained on a TCGA Head & Neck Squamous Cell Carcinoma (HNSC) cohort.

⚠️ Read this first: what's real vs. what's scoped

The brief describes 4 data modalities (images, genome, transcriptome, epigenome). Only 3 were actually provided — no .svs histopathology slides were in the uploaded dataset, only 7 tabular files:

File	Modality	Rows
clinical.txt	Clinical	309 rows → 82 patients
exposure.txt	Risk factors (smoking/alcohol)	114 rows → 82 patients
follow_up.txt	Longitudinal follow-up	279 rows → 82 patients
pathology_detail.txt	Pathology detail	175 rows → 82 patients
mutations.txt	Genome (MAF-style)	13,180 mutation calls, 82 tumor samples
transcriptomics.txt	Transcriptome (RNA-seq)	20,502 genes x 92 samples
methylation.txt	Epigenome (methylation)	20,114 probes x 96 samples

This is also one cancer type, not a cancer-type-agnostic dataset (cases.disease_type = "Squamous Cell Neoplasms", primary sites = tongue, tonsil, floor of mouth, gum, palate, lip → classic HNSC). That reshapes the 4 tasks into their honest, data-supported form:

Brief's task	What we actually built	Why
Cancer vs Non-Cancer	Tumor vs Normal tissue classification	Every sample is HNSC; but ~10-14 patients have a matched normal tissue sample (barcode suffix -11 vs -01), so tumor-vs-normal is the real binary distinction in the data.
Stage I-IV multiclass	Early (I/II) vs Late (III/IV) binary	Only 15/27/25/108/2 patients per stage — 4-way classification on ~80 patients would be unstable/meaningless. Binarizing gives usable class sizes.
Risk factor prediction	Smoking status from molecular signature (omics only — other exposure fields are explicitly excluded as features, see leakage note below)	exposure.txt has real tobacco_smoking_status / alcohol_history fields; used as labels, not inputs.
Advanced insights	Cox survival risk score, SHAP-driven biomarkers, patient similarity, subtype discovery	All real, computed on the same cohort.
Histopathology imaging	Not trained — stubbed	No .svs files were uploaded. See src/image_pipeline_stub/README.md for exactly how to plug it in once slide data exists.

A data-leakage bug was caught and fixed during development: predicting smoking status using other exposure-table fields (e.g. pack_years_smoked) would trivially "solve" the task without touching biology. main.py now explicitly excludes all exposures.* fields from Task 3's inputs, and all diagnoses.ajcc* (T/N/M) fields from Task 2's inputs (since stage is literally derived from those). Metrics below are honest, post-fix numbers.

Results (5-fold stratified cross-validation, not in-sample)

Task	Best model	Accuracy	Macro F1	Cohen's κ	ROC-AUC	n
1. Tumor vs Normal	Logistic Regression	0.977	0.894	0.788	1.000	88 samples
2. Stage (Early vs Late)	LightGBM	0.686	0.626	0.252	0.627	70 patients
3. Smoking status (molecular)	Logistic Regression	0.835	0.751	0.503	0.765	79 patients
4. Survival (Cox)	Cox PH (PCA + clinical)	c-index = 0.721	—	—	—	82 patients

Task 2's weaker score (κ=0.25, ROC-AUC=0.63, close to random) is the honest result of only 70 patients spread across two imbalanced classes with heavy omics dimensionality — flagged here rather than hidden. Task 1 and Task 3 are both strong and clinically plausible: tumor-vs-normal has a huge, well-known expression signature, and there is real published literature on smoking leaving a DNA-methylation signature (predicting it purely from omics, with all direct exposure fields excluded, at ROC-AUC 0.77 is a genuinely interesting result for a hackathon).

Task 4 discovered 2 molecular subtypes (silhouette = 0.246, modest but real cluster separation) — see the Analytics page for the subtype-vs-stage cross-tab.

Project structure

project/
├── data/
│   ├── raw/                  # your 7 input files
│   └── processed/            # patient/sample-level modeling tables (generated)
├── src/
│   ├── preprocessing/        # clinical/exposure/follow-up/pathology loader, genome + omics loaders
│   ├── training/              # classification.py (tasks 1-3), survival.py (task 4)
│   ├── explainability/       # SHAP (tree + linear models)
│   ├── inference/            # predict.py — CLI scoring
│   ├── utils/                # TCGA barcode parsing
│   ├── image_pipeline_stub/  # documented, inactive — see its README
│   └── build_dataset.py      # step 1: raw -> data/processed/*.csv
├── models/                   # saved best model per task (.joblib, generated)
├── reports/                  # evaluation JSON + SHAP figures (generated)
├── streamlit_app/            # multi-page dashboard
├── main.py                   # step 2: train + evaluate + explain all 4 tasks
├── config.yaml
├── requirements.txt
└── Dockerfile

Running it

pip install -r requirements.txt --break-system-packages   # or use a venv

# Step 1: raw TSVs -> processed modeling tables
python src/build_dataset.py

# Step 2: train, evaluate, explain all 4 tasks
python main.py

# Step 3: launch the app
streamlit run streamlit_app/app.py

Re-running build_dataset.py + main.py after dropping new files into data/raw/ (matching the same schema) will re-train everything from scratch — nothing is hardcoded to this specific 82-patient cohort.

CLI single-patient scoring, without the app:

python src/inference/predict.py --task task2_stage --id TCGA-CR-7392

Streamlit app pages

• Dashboard — cohort overview, KPIs, stage/smoking/vital-status breakdowns.
• Upload Data — schema-validating upload for MAF / transcriptome / methylation files (histopathology tab is disabled — see note above).
• Prediction — pick a patient/sample, see the model's prediction vs. ground truth, class probabilities, and a full-cohort batch prediction table.
• Explainability — SHAP importance plots per task; Cox hazard ratios and PCA-loading biomarkers for the survival model.
• Analytics — survival risk distribution, patient-similarity lookup, discovered-subtype breakdown.
• Leaderboard — every model tried on every task, cross-validated, sortable.

Deployment

Streamlit Community Cloud (simplest)

1. Push this repo to GitHub.
2. On https://share.streamlit.io, point at streamlit_app/app.py, and set the requirements file to requirements.txt.
3. Run python src/build_dataset.py && python main.py once locally first and commit data/processed/, models/, and reports/ — the app reads pre-computed artifacts rather than training on every page load.

Docker (Render / Railway / any container host)

docker build -t cancer-omics-platform .
docker run -p 8501:8501 cancer-omics-platform

Local machine

streamlit run streamlit_app/app.py

Methodology notes (for the judges)

• Small-n, high-p handling: every classification pipeline is StandardScaler → SelectKBest(f_classif, k=50) → model, evaluated with Stratified K-Fold (not a single holdout, which would be unstable at n=70-88) so the reported numbers are honest out-of-fold performance.
• Survival modeling: with 82 patients and ~4,000 raw omics features, a Cox model can't be fit directly (p >> n). We PCA-reduce the omics block to 10 components, combine with age + mutation burden, and fit a penalized Cox model — then trace the top PCA loadings back to raw genes/probes for interpretable "biomarker" reporting.
• Patient similarity / subtypes are computed in the same PCA space via cosine similarity and KMeans (silhouette-selected k), respectively.
• Leakage guards are explicit and documented in main.py / config.yaml (leakage_prefixes) rather than implicit — this is the kind of thing that's easy to accidentally get "too good" results on.

Recent changes

A full, versioned history lives in CHANGELOG.md. Highlights of the most recent iteration:

• Redesigned dashboard UI — a modern, professional healthcare look applied consistently across every page, driven by a single shared theme in streamlit_app/ui/theme.py (clinical palette, shared Plotly template, hero headers, section titles, metric cards, styled tabs, and card containers).
• Sidebar credits — a credit block pinned to the bottom of the sidebar on every page (Built by Salman Khan & Team · NeoHack GIKI Edition 2025 · in collaboration with Precision Medicine Lab).
• scikit-learn compatibility fix — resolved AttributeError: 'LogisticRegression' object has no attribute 'multi_class' on the Prediction page (caused by loading models pickled with scikit-learn 1.8.0 under runtime 1.7.2). The model loader now backfills the missing attribute at load time so predict_proba works across tasks.
• Project governance docs — added CONTRIBUTING.md, CODE_OF_CONDUCT.md, CHANGELOG.md, and .gitignore.

Contributing

Contributions are welcome. Please read CONTRIBUTING.md for setup, coding standards, UI/theming conventions, and pull-request guidelines, and follow our Code of Conduct.

License

Released under the MIT License.