GNN-Based Prediction of Drug-Drug Interactions Using Pretrained SMILES Embeddings

Problem Statement

Due to the small numbers of participants in clinical studies, many DDIs remain undetected. Similarity-based computational methods struggle worth complex interactions. Traditional Graph Neural Networks (GNNs) do not prioritise pharmacologically relevant neighbours. The research gap consists of a lack of frameworks that combine molecular representation learning with graph structural context in a focused manner. Wet-lab DDI detection is time-consuming, costly, and limited in scale. A vast number of possible drug combinations makes exhaustive testing impossible. Use of SMILES-based molecular embeddings + Graph Attention Networks (GAT) enables advanced pattern recognition.

Architecture Diagram

DrugBank SMILES Dataset

We use two medication interaction databases that are considered the best: DrugBank and BioSNAP. DrugBank provides carefully structured pharmacological data andinformation regarding interactions between drugs, whileBioSNAP delivers additional drug-drug interaction insights obtained from chemical and clinical sources. To ensure consistency and chemical validity, we retrieve the SMILES (Simplified Molecular Input Line Entry System) strings for each drug and validate them using RDKit. Invalid or incomplete SMILES entries are filtered out. The remaining validated drugs are retained and form the node set for graph construction. Duplicate and noninteracting entries are removed to yield a high-confidence, connected interaction graph

SMILES Data From DrugBank	Count
Total Drugs in Dataset	12227
Total SMILES in Dataset	12227
Valid SMILES in Dataset	11574
Removed Drugs in Dataframe	653

Graph Construction

A heterogeneous, undirected graph is constructed where each node corresponds to a unique drug, and edges represent known drug–drug interactions from the merged DrugBank and BioSNAP datasets. We only keep medications that have valid embeddings and validated interaction edges. The final graph is converted into a format compatible with PyTorch Geometric, including the edge list and feature matrix. Every node consequently contains a vectorized embedding, while each edge signifies an interaction that needs to be predicted or validated. To pre-process the dataset for link prediction, we use Random Link Split, a process that splits the edges into training, validation, and test sets in a balanced way, representing positive and negative samples. Further, negative sampling is used in the training process to mimic unknown (non-interacting) drug pairs in order to increase the model’s discriminative ability.

Dataset	BioSnap	DrugBank	CombinedGraph
Number of Nodes	1514	1706	1756
Number of Edges	48514	191402	198020
Is Undirected	True	True	True
Isolated Nodes	False	False	False
Isolated Loops	False	False	False
Average Node Degree	64.087	224.38	225.53

Results