🧬 BioGNN: Drug–Target Interaction Link Prediction

GNN-based target prioritization for AI-driven drug discovery

---

🔬 Overview

This repository demonstrates an end-to-end Graph Neural Network (GNN) pipeline for drug–target interaction (DTI) link prediction, built as a portfolio project for AI-driven target discovery and translational research.

It is intentionally minimal, reproducible, and runnable from the terminal, while reflecting real industry patterns used in computational biology and drug discovery.

This project complements an upstream LLM-based hypothesis generator:

LLM → biological relation extraction → candidate targets ↓ GNN → link prediction → prioritization

---

🚀 What This Project Demonstrates

✅ Technical Skills

• PyTorch Geometric (GCN-based GNN)
• Graph construction for biological networks
• End-to-end training & inference pipeline
• Model checkpointing and reuse
• CLI-driven reproducibility

✅ Drug Discovery Relevance

• Drug–target interaction modeling
• Latent biological representation learning
• Target prioritization via learned embeddings
• Ready to extend to ChEMBL / DrugBank / CRISPR / multi-omics graphs

---

🧠 Model Architecture

Node features (x) ↓ GCNConv (in_dim → hidden_dim) ↓ ReLU ↓ GCNConv (hidden_dim → hidden_dim) ↓ Node embeddings (latent biological space) ↓ Link score (dot product / classifier)

• Nodes: drugs, proteins (demo-scale)
• Edges: known or hypothesized interactions
• Output: link score representing interaction likelihood

---

📂 Project Structure

biognn-dti-link-prediction/
├── src/biognn/
│   ├── data.py        # graph construction
│   ├── model.py       # DTI_GNN (GCN)
│   ├── train.py       # training + checkpoint save
│   └── infer.py       # inference + link scoring
├── scripts/
│   ├── 01_build_graph.py
│   ├── 02_train.sh
│   └── 03_infer.sh
├── outputs/
│   └── gnn_model.pt   # trained model checkpoint
└── README.md

---

⚙️ How to Run (Reproducible Demo)

1) Build Graph

python scripts/01_build_graph.py

2) Train Model

bash scripts/02_train.sh

Example output:

[INFO] Building demo graph...
[INFO] Graph: num_nodes=4, num_edges=3
[INFO] Starting training...
Epoch 001 | Loss: ...
...
[DONE] Model saved to outputs/gnn_model.pt

3) Run Inference

bash scripts/03_infer.sh

Example output:

[INFO] Loading demo graph...
[INFO] Graph: num_nodes=4, num_edges=3
[INFO] Embeddings shape: (4, 32)
[RESULT] Example link score
  node_i=0  node_j=3  score=0.2740

---

🔍 Interpretation

• Node embeddings represent learned biological states in a latent space.

• Link score represents a predicted interaction likelihood between two nodes.

• In real applications, this supports:

  • target prioritization
  • drug repurposing hypotheses
  • mechanism exploration
  • experimental design guidance

---

🔄 Extension Ideas (Real-World Ready)

This pipeline is designed to scale to:

• ChEMBL / DrugBank networks
• PPI graphs and pathway graphs
• CRISPR perturbation graphs
• Multi-omics feature integration (transcriptomics, proteomics)
• LLM-generated hypothesis edges
• GNN + LLM hybrid reasoning for target discovery

---

👤 Author

Dohoon Kim Senior Computational Biologist / Data Scientist Focus: AI for drug discovery, target identification, and translational biology

---

⭐ Why This Matters

This repository demonstrates the ability to:

• translate biology into graphs
• apply GNNs to discovery problems
• build reproducible training/inference pipelines
• connect LLM-derived hypotheses to mechanistic graph modeling

These are core skills required for AI Computational Biologist roles.