Gene Biomarker Discovery & Drug Repurposing for Pancreatic Cancer (PDAC)

This repository contains the code and experimental workflow for identifying robust and interpretable gene biomarkers in Pancreatic Ductal Adenocarcinoma (PDAC) using LSTM-based deep learning combined with Explainable AI (XAI) methods.

---

🚀 Key Highlights

Component	Details
Dataset	Microarray gene expression dataset (NCBI GEO: GSE183795)
Feature Selection	ANOVA F-test (SelectKBest)
Model	Long Short-Term Memory (LSTM) for gene dependency learning
Interpretability	SHAP & LIME
Evaluation	5-fold stratified CV
Identified Biomarkers	KRT19, CTSE, LAMC2, ANXA2

---

📊 Best Model Performance (K = 20 Features)

Metric	Score
Accuracy	80.47%
Precision	86.34%
Sensitivity (Recall)	83.49%
F1-Score	83.33%

These results demonstrate that a minimal gene signature (20 genes) can effectively classify PDAC vs. Normal samples.

---

🔍 Explainability

Two complementary XAI techniques were applied to ensure model transparency:

Method	Purpose
SHAP	Global feature importance across all samples
LIME	Local decision explanation per subject

Both methods consistently highlighted the same 4 key biomarkers, reinforcing biological reliability.

---

📁 Repository Structure

notebooks/
  train_transformer.ipynb   TabTransformer model for PDAC classification
  train_lstm.ipynb          LSTM model (gene-dependency learning) + SHAP/LIME
  train_mlp.ipynb           MLP baseline
models/
  tabular_transformer_pancreas.pt   Trained TabTransformer weights
  tabular_lstm_pancreas.pt          Trained LSTM weights
  tabular_mlp_pancreas.pt           Trained MLP weights
README.md

🚀 Usage

Open a training notebook in Jupyter or Colab and run all cells:

• notebooks/train_lstm.ipynb — the primary model (LSTM + SHAP/LIME explainability).
• notebooks/train_transformer.ipynb and notebooks/train_mlp.ipynb — comparison baselines.

Each notebook trains on the GSE183795 expression data and saves weights to a .pt file (the pre-trained weights in models/ are included for reference). A GPU runtime is recommended.

Note: notebooks save weights to the working directory by default — point the save/load paths at models/ if you want to regenerate them in place.