Onco-Logic: An Integrated AI Suite for Precision Oncology

---

Overview

Onco-Logic is a comprehensive, multi-modal decision support ecosystem designed to transform cancer care by unifying fragmented patient data. The suite leverages advanced AI and machine learning to provide clinicians and researchers with a holistic understanding of each patient's disease, enabling a new frontier in precision oncology.

The project is built on three synergistic pillars, each targeting a critical data modality in cancer research and clinical practice:

• Pathfinder-NLP: An NLP pipeline that extracts and standardizes critical information from unstructured, free-text pathology reports using a fine-tuned ClinicalBERT model. It transforms narrative documents into structured, computable data for large-scale analysis and clinical decision support.
• Geno-Classify: A machine learning framework that analyzes complex gene expression profiles to perform precise, automated cancer subtyping. It addresses the "high-dimension, low-sample-size" challenge to guide targeted therapies and uncover biological insights.
• Pred-Nosis: A predictive modeling tool that forecasts breast cancer patient survival status (Alive/Deceased) and duration. It employs advanced data resampling techniques (SMOTE+ENN) to handle severe class imbalance and achieve high-accuracy predictions.

This repository contains the complete Streamlit application, which provides an interactive interface for exploring datasets, evaluating models, and utilizing the predictive capabilities of each pillar.

---

Tech Stack

• Backend & ML: Python 3.9+, PyTorch, Scikit-learn, Transformers (Hugging Face), SHAP, UMAP, Imbalanced-learn
• Frontend: Streamlit
• Data Manipulation: Pandas, NumPy
• Plotting & Visualization: Matplotlib, Seaborn
• NLP: SpaCy, NLTK
• LLM Integration: OpenRouter API for generating prognostic summaries
• Environment Management: dotenv for managing API keys

---

Features & Application Pages

The Onco-Logic suite is delivered as an interactive Streamlit application with the following key modules:

1. NLPathology (Pathfinder-NLP)

An end-to-end NLP pipeline for processing and analyzing oncological pathology reports.

• Interactive Report Processor: Upload or paste raw pathology text to automatically extract structured data, including cancer type and TNM staging, using a fine-tuned ClinicalBERT model.
• AI-Powered Summaries: Integrates with the OpenRouter API to generate professional, medically accurate prognostic summaries based on the extracted clinical information.
• Unsupervised Text Exploration: Utilizes TF-IDF vectorization with PCA, t-SNE, and UMAP to visualize high-dimensional text data and reveal inherent clustering of reports by cancer type.
• N-gram Frequency Analysis: Identifies the most common bi-grams and tri-grams in the text corpus to uncover key clinical terminology and patterns.
• Named Entity Recognition (NER): A demonstration using SpaCy to automatically highlight and classify clinical entities within pathology reports.

2. CancerGenetics (Geno-Classify)

A comprehensive toolkit for cancer subtype classification based on gene expression data.

• Clustering Analysis: Employs K-Means clustering on data processed through various feature engineering pipelines (e.g., Variance Thresholding, UMAP, Log Transformation) to evaluate unsupervised subtype separation.
• Model Evaluation: Trains and evaluates multiple classifiers (Decision Tree, Random Forest, SGD Classifier) using stratified cross-validation to predict cancer subtypes with high accuracy.
• SHAP Feature Importance: Uses SHAP (SHapley Additive exPlanations) to interpret model predictions and identify the most influential genes driving classification, bridging the gap between prediction and biological insight.
• Clinical Predictor Tool: An interactive tool to predict cancer subtypes based on user-inputted gene expression values, leveraging the best-performing trained models.

3. BreastCancerSurvival (Pred-Nosis)

A robust application for predicting breast cancer patient outcomes using structured clinical data.

• Multi-Class Survival Prediction: Forecasts survival duration across multiple time-window categories (e.g., <1 year, 1-2 years) with high accuracy.
• Advanced Resampling: Implements a state-of-the-art hybrid resampling technique (SMOTE+ENN) to address severe class imbalance in survival data, boosting Random Forest model accuracy to 95.9%.
• Model Bake-off: Systematically trains, evaluates, and compares multiple models (Random Forest, XGBoost, CatBoost, LightGBM) on raw, oversampled, and cleaned datasets.
• Interactive Risk Predictor: Allows users to input patient characteristics and receive a survival risk prediction, complete with model confidence scores.

---

Project Structure

Onco-Logic/
├── App.py                    # Main Streamlit application launcher
├── requirements.txt          # Project dependencies
├── .env                      # Environment variables for API keys
├── .env.example              # Example environment file
├── .gitignore                # Specifies files and directories for Git to ignore
├── .gitattributes            # Configures Git Large File Storage (LFS) for large files
├── Data/                     # Directory for datasets
├── models/                   # Directory for saved models
├── pages/
│   ├── BreastCancerSurvival.py # Streamlit page for the "Pred-Nosis" module
│   ├── CancerGenetics.py       # Streamlit page for the "Geno-Classify" module
│   ├── NLPathology.py          # Streamlit page for the "Pathfinder-NLP" module
│   └── Secondary Endpoint/
│       └── BreastCancerStatus.py # Additional analysis page
├── resources/                # Directory for additional resources (for cancer genetics models)
└── scripts/
    ├── cancer_trainer.py       # Script for training the cancer type classification model
    ├── stage_trainer.py        # Script for training the TNM staging models
    ├── llm_api.py              # Module for interacting with the OpenRouter LLM API
    ├── custom_transformers.py  # Custom scikit-learn transformers for ML pipelines
    └── __pycache__/            # Cache directory for compiled Python files
        ├── custom_transformers.cpython-313.pyc
        └── llm_api.cpython-313.pyc

---

Quick Start (Local)

1. Prerequisites

• Python 3.9+ and pip
• Git and Git LFS for handling large model files

2. Clone & Install

git clone <this-repo-url>
cd Onco-Logic
pip install -r requirements.txt

3. Configure Environment

Copy the example environment file and add your OpenRouter API key to enable the AI-powered prognosis summary feature in the NLPathology tool.

cp .env.example .env
# Then edit .env and add your key

Required keys:

• OPENROUTER_API_KEY: Your API key from OpenRouter.ai.

4.. Datasets and Models

This repository includes the datasets and pre-trained models required to run the application.

• Datasets: The following CSV files are located in the Data/ directory:
  • cancer_subtype_data.csv
  • cancer_subtype_labels.csv
  • Breast_Cancer.csv
  • TCGA_M01_patients.csv
  • TCGA_N03_patients.csv
  • TCGA_patient_to_cancer_type.csv
  • TCGA_Reports.csv
  • TCGA_T14_patients.csv
• Models: The pre-trained NLP models are included in the models/ directory. You can also retrain them using the provided scripts:
  • Run scripts/cancer_trainer.py to train the cancer type classification model.
  • Run scripts/stage_trainer.py --stage all to train the T, N, and M staging models.

5. Run the App

Once the environment is configured and datasets are in place, run the Streamlit application:

streamlit run App.py

The application will be available at http://localhost:8501. Use the sidebar to navigate between the different modules of the Onco-Logic suite.