🧬 Multimodal GAT for Drug–Drug Interaction Prediction

A deep learning framework for predicting Drug–Drug Interactions (DDIs) using multimodal Transformer embeddings and Graph Attention Networks (GAT).

---

Problem Statement

Drug–Drug Interactions (DDIs) are often undetected due to limited clinical trials and the vast number of possible drug combinations. Traditional similarity-based methods fail to capture complex pharmacological relationships, while standard Graph Neural Networks (GNNs) do not prioritize biologically relevant neighbors. This work addresses the gap by integrating:

• Molecular representations (SMILES)
• Semantic drug descriptions
• Graph structural context (GAT)

to enable accurate and scalable DDI prediction.

---

Architecture

The proposed framework consists of:

1. Multimodal Encoding

   • SMILES and descriptions encoded using Transformer models (T5, SBERT, ChemBERTa)

2. Fusion Layer

   • Concatenation + projection to unified embeddings

3. Graph Attention Network

   • Learns drug interaction patterns via attention-based aggregation

4. Link Prediction

   • MLP + Sigmoid for interaction probability

---

Dataset

We use two widely accepted biomedical datasets:

• DrugBank → Drug attributes (SMILES + descriptions)
• BioSNAP → Drug–drug interaction network

SMILES Data Summary

Metric	Count
Total Drugs	12,227
Total SMILES	12,227
Valid SMILES	11,574
Removed Invalid Entries	653

SMILES strings are validated using RDKit to ensure chemical correctness.

---

Graph Construction

A heterogeneous undirected graph is constructed where:

• Nodes → Drugs
• Edges → Known drug interactions

Graph Statistics

Dataset	BioSNAP	DrugBank	Combined
Nodes	1514	1706	1756
Edges	48,514	191,402	198,020
Undirected	Yes	Yes	Yes
Isolated Nodes	No	No	No
Avg Degree	64.08	224.38	225.53

Preprocessing Steps

• SMILES validation (RDKit)
• Removal of invalid entries
• Graph construction using PyTorch Geometric
• Random Link Split (train/val/test)
• Negative sampling for non-interacting pairs

---

Methodology

1. Multimodal Embedding

• Transformer-based encoders generate embeddings:
  • Structural (SMILES)
  • Semantic (Descriptions)

2. Fusion

• Concatenation + Linear Projection
• Produces unified drug representation

3. Graph Learning

• Two-layer Graph Attention Network (GAT)
• Learns relational dependencies

4. Link Prediction

• Pairwise feature construction:
• MLP decoder → Sigmoid output

---

Results

ROC Curve

Precision-Recall Curve

---

Performance Summary

Model	Representation	AUC	AUPR
T5	SMILES	0.872	0.861
SBERT	SMILES	0.888	0.870
ChemBERTa	SMILES	0.829	0.787
T5	Description	0.847	0.818
SBERT	Description	0.931	0.925
ChemBERTa	Description	0.817	0.764
T5	Fusion	0.888	0.880
SBERT	Fusion	0.945	0.943
ChemBERTa	Fusion	0.866	0.831

Best Model: SBERT Fusion
Key Insight: Multimodal fusion significantly improves performance

---

Tech Stack

• Python
• PyTorch & PyTorch Geometric
• HuggingFace Transformers
• RDKit
• Scikit-learn

---