Pan Cancer Classification Pipeline

Multi-class, three-class, and binary classification of cancer types using RNA-seq gene expression data (TCGA Pan-Cancer dataset). All three tasks run from a single unified pipeline.

Tasks

Task	Classes	Target Column
multiclass	24 TCGA cancer types + Normal	Class
three_class	LIHC / Others / Normal	LC_labels
binary	LIHC / Others	LC_labels

Models

Random Forest, Extra Trees, XGBoost — each trained as original and Optuna-tuned.

Structure

cancer-classification/
├── data/
│   ├── train.csv  
│   └── test.csv 
├── models/
│   ├── multiclass/
│   │   ├── saved_models/
│   │   ├── confusion_matrices/
│   │   └── classwise_metrics/
│   ├── three_class/
│   └── binary/
├── results/
│   ├── multiclass/
│   ├── three_class/
│   └── binary/
├── config.py 			# paths, target mappings, task configs
├── evaluate.py			# metric functions
├── train.py 			# unified pipeline
├── requirements.txt
└── README.md

---

Dependencies

Package	Version	Purpose
pycaret[full]	3.3.2	ML pipeline (training, tuning, evaluation)
scikit-learn	1.5.2	Metrics — confusion matrix, AUC, MCC
pandas	2.2.3	Data loading and results saving
numpy	1.26.4	Numerical computations
optuna	4.1.0	Hyperparameter tuning (used inside PyCaret)
joblib	1.4.2	Model serialization

Note: pycaret[full] installs most of the above automatically.
Pinned versions ensure reproducibility across machines.

---

Setup

# You can place your data files and run
cp your_train.csv data/train.csv
cp your_test.csv  data/test.csv

Usage

# Run a specific task
python train.py --task multiclass  --optuna_itr 15       # defaults is 10
python train.py --task three_class
python train.py --task binary

# Run all three tasks in sequence
python train.py --task all

Outputs (per task)

File	Description
results/<task>/combined_results_original.csv	CV metrics — original models
results/<task>/combined_results_tuned.csv	CV metrics — tuned models
results/<task>/all_results.csv	Train + test results combined
results/<task>/all_classwise_metrics.csv	Per-class Sensitivity, Specificity, Precision, F1, MCC, AUC
models/<task>/confusion_matrices/	One CSV per model variant
models/<task>/saved_models/	.pkl files (original + tuned)