Business Context

OncoMap Bio is building a clinical decision-support workflow that uses next-generation sequencing (NGS) tumor profiles to help researchers distinguish cancer subtypes and prioritize downstream analysis. You need to train a machine learning model that predicts cancer type from NGS-derived features while remaining interpretable enough for translational research teams.

Dataset

You are given a retrospective dataset of tumor samples collected from multiple hospital partners.

• Size: 12,500 tumor samples, 186 features
• Target: Multiclass label for 5 cancer types: breast, lung, colorectal, melanoma, ovarian
• Class balance: Moderately imbalanced; largest class 31%, smallest class 11%
• Missing data: ~8% missing in clinical covariates and ~3% missing in copy number features

Success Criteria

A good solution should achieve strong multiclass discrimination with macro F1 >= 0.78 and one-vs-rest ROC-AUC >= 0.90 on a held-out test set. The team also wants feature importance outputs to explain which genomic signals drive predictions.

Constraints

• Inference will run in batch on up to 20K samples per week
• Researchers need interpretable feature importance, not a black-box-only solution
• Training must avoid leakage across hospital sites and sequencing batches
• Budget favors classical ML over large deep learning models

Deliverables

1. Build a multiclass classification pipeline for cancer type prediction from NGS features.
2. Explain why NGS-derived features are informative for cancer research and classification.
3. Handle missing values, mixed feature types, and class imbalance appropriately.
4. Evaluate the model with suitable multiclass metrics and justify threshold-free metrics.
5. Provide feature importance or coefficient-based interpretation for the top predictive genomic features.