CBioHackathon

Protein-Protein Interaction (PPI) link prediction on the STRING database network. We benchmark classical graph heuristics against GNN-based methods across two task types: edge prediction (transductive) and node prediction (inductive / cold-start).

Dataset

File	Description
string_interaction_physical.tsv	Full STRING network (~4 300 edges, 354 proteins)
string_interactions_short.tsv	Subset for fast iteration (~300 edges, 175 proteins)
string_protein_sequences.fa	FASTA protein sequences (for ESM-2 embeddings)
protein_embeddings.pkl	Pre-computed ESM-2 (650M) embeddings

Methods

Edge Prediction (Transductive)

Method	Script	Key Idea
Random Baseline	random_baseline.py	Uniform random scores -- lower bound
Common Neighbors	CN_baseline/	Weighted CN scoring with threshold tuning
Random Walk w/ Restart	markov_baseline.py	RWR affinity matrix with alpha tuning
Sequence-only MLP	pred_by_seq_baseline.py	Concatenated ESM-2 pair embeddings, MLP classifier (PyTorch Lightning)

Node Prediction (Inductive / Cold-Start)

Method	Script	Key Idea
Adamic-Adar + Sequence	adamic_adar_sequence.py	Sequence similarity creates virtual neighbors; score via AA index
VGAE + MLP (GCN)	gnn_lightning_metrics.py	VGAE learns structure; MLP maps ESM to latent for unseen nodes
VGAE + MLP (GAT)	ALON_BEST_gat_lightning_metrics.py	Same pipeline with multi-head GAT encoder
Node-deletion GNN	node_deletion_gnn.py	VGAE on partial graph; evaluate reconstruction of deleted nodes

Supporting Code

File	Purpose
utils.py	Graph loading, train/val/test splits (edge, vertex, semi-inductive), metrics, ROC plotting
extract_proteins_representations.py	Extract ESM-2 embeddings from FASTA to pickle
alon_files/	VGAE experiments, topology analysis, CN vs GNN comparison scripts

Quick Start

# Setup
python -m venv venv && source venv/bin/activate
pip install -r requirements.txt

# Generate protein embeddings (one-time, ~10 min)
python extract_proteins_representations.py --fasta string_protein_sequences.fa

# Run baselines
python random_baseline.py
python markov_baseline.py

# Run sequence-based classifier
python pred_by_seq_baseline.py

# Run cold-start node prediction
python adamic_adar_sequence.py
python gnn_lightning_metrics.py

Requirements

torch>=2.0.0
fair-esm>=2.0.0
torch-geometric
pytorch-lightning
torchmetrics
scikit-learn
numpy>=1.24.0
pandas>=2.0.0
networkx>=3.0
matplotlib>=3.7.0
seaborn>=0.12.0

Project Structure

CBioHackathon/
├── utils.py                          # Shared graph utilities, splits and metrics
├── extract_proteins_representations.py # ESM-2 embedding extraction
├── random_baseline.py                # Random baseline
├── markov_baseline.py                # Random Walk with Restart
├── pred_by_seq_baseline.py           # Sequence-only MLP (Lightning)
├── adamic_adar_sequence.py           # AA + sequence for node prediction
├── gnn_lightning_metrics.py          # VGAE cold-start (GCN encoder)
├── ALON_BEST_gat_lightning_metrics.py # VGAE cold-start (GAT encoder)
├── ALON_BEST_gnn_lightning_metrics.py # VGAE cold-start (GCN, with t-SNE viz)
├── node_deletion_gnn.py              # Inductive node-deletion experiment
├── CN_baseline/                      # Common Neighbors baseline and evaluation
├── alon_files/                       # VGAE experiments and topology analysis
├── Report/                           # Report guidelines
├── string_interaction_physical.tsv   # Full dataset
├── string_interactions_short.tsv     # Small dataset
├── string_protein_sequences.fa       # Protein sequences
├── protein_embeddings.pkl            # Pre-computed ESM-2 embeddings
└── requirements.txt