Heterogeneous Graph Neural Network-based Drug-Target Interaction Prediction: A Computational Platform to Accelerate Drug Discovery

Core Project Information

• Original Author/Maintainer: Babakmamnoon
• Source Platform: GitHub
• Original Title: Drug-Target-Interaction-Prediction-on-Heterogeneous-Graphs-using-Graph-Neural-Networks
• Original Link: https://github.com/Babakmamnoon/Drug-Target-Interaction-Prediction-on-Heterogeneous-Graphs-using-Graph-Neural-Networks
• Publication Time: June 2026
• Colab Notebook: https://colab.research.google.com/drive/18GwTTIiTozvVw4e1jbSVjC4VJWPB4Rlr

Core Insights

This project provides a complete industrial-grade workflow for drug-target interaction (DTI) prediction using heterogeneous graph neural networks (GNNs) on the BioSNAP-DTI benchmark dataset. By combining molecular graph characterization with a 1D CNN protein encoder, it achieves high-precision binary classification of DTIs and reaches state-of-the-art performance.

Guide to Subsequent Floors

Subsequent floors will sequentially cover research background, dataset details, methodology, experimental results, technical highlights and application prospects, and project conclusion.

Predicting the physical binding between small-molecule drugs and protein targets (DTI prediction) is a fundamental problem in computational drug discovery. Experimental characterization of DTIs is costly (≈$1.8 billion per new molecular entity approval) and time-consuming (9-12 years). High-throughput computational screening can accelerate this process, but traditional methods rely on feature engineering (e.g., molecular fingerprints, handcrafted protein features), which fail to capture structural hierarchies and have limited information content.

Graph neural networks (GNNs) represent drugs as molecular graphs (atoms as nodes, chemical bonds as edges) and learn structural representations end-to-end. Combined with sequence-based protein encoders and explicit interaction modeling layers, GNNs perform excellently on DTI benchmarks.

BioSNAP-DTI Dataset

BioSNAP-DTI is a widely used binary DTI classification benchmark constructed and preprocessed by the Stanford SNAP Lab.

Dataset Statistics

Attribute	Value
Total DTI Pairs	~27,462
Unique Drugs	4,510
Unique Protein Targets	2,181
Positive Interactions	~13,830 (50%)
Negative Interactions	~13,632 (50%)
Drug Representation	SMILES
Protein Representation	Amino Acid Sequence
Data Source	DrugBank 5.0, Stanford SNAP MINER
Standard Split	Train/Validation/Test ≈70/10/20%

Data Cleaning Steps

1. Remove rows with missing SMILES, protein sequences, or labels
2. Validate and normalize SMILES via RDKit
3. Validate protein sequences (only accept standard 20 amino acids)
4. Limit protein length to 1200 amino acids (covers >95% of data, prevents memory overflow)

Drug Characterization: Molecular Graph

Each drug SMILES is converted into a PyTorch Geometric Data object, using 27-dimensional atomic features (atom type, hybridization, aromaticity, etc.) and 6-dimensional bond features (bond type, conjugation, ring membership).

Protein Characterization: 1D CNN Encoding

Protein sequences are encoded into integer tensors, processed via three parallel 1D CNN branches (kernel sizes: 3,7,11), and adaptive max-pooling yields 256-dimensional embeddings to capture local to medium-range features.

DTI-GNN Model Architecture

Model Flow:

1. Drug SMILES → Input Projection → ResGCN Blocks ×3 → Global Average + Max Pooling → 256D Drug Embedding
2. Protein Sequence → Embedding Layer → Parallel CNN → Adaptive Pooling →256D Protein Embedding
3. Bilinear Attention Module (models interactions) → MLP Classifier (256→128→64→2) → Binary Prediction

Key Design Decisions

• Residual connections stabilize deep gradients
• Dual pooling strategy preserves atomic signals
• Bilinear attention explicitly models drug-target interactions
• Multi-scale CNN kernels capture protein motifs

Training Strategy

Hyperparameter	Value
Optimizer	Adam
Learning Rate	5×10⁻⁴
Weight Decay	1×10⁻⁵
Batch Size	64
Max Epochs	50
Learning Rate Scheduler	ReduceLROnPlateau
Early Stopping	Patience=10 epochs
Loss Function	Cross Entropy
Gradient Clipping	Norm=1.0
Dropout	0.3

Test Set Performance

Metric	Value
AUROC	0.951
AUPRC	0.948
Accuracy	0.892
F1 Score	0.891
Precision	0.888
Recall	0.894
MCC	0.784

5-Fold Cross-Validation

Average AUROC of 0.950 with a standard deviation of 0.002, indicating stable model generalization.

Benchmark Comparison

The project's DTI-GNN outperforms previous state-of-the-art methods on BioSNAP-DTI:

Method	Year	AUROC	AUPRC
DrugBAN	2023	0.948	0.945
DTI-GNN	2024	0.951	0.948

Visualization Tools

Provides radar charts of metrics, ROC/PRC curves, confusion matrices, molecule-level prediction visualizations, etc.

Technical Highlights

1. Heterogeneous Graph Learning: Handles heterogeneous data (molecular graphs + protein sequences) and maps them to a unified embedding space.
2. Multi-scale Feature Extraction: Parallel CNN kernels capture protein structural information at different levels.
3. Bilinear Attention: Explicitly models drug-target interactions, outperforming simple concatenation.
4. Industrial-grade Practice: Includes complete workflows like data cleaning, early stopping, cross-validation, etc.

Application Prospects

• Drug Repurposing: Predict new targets for existing drugs
• Side Effect Prediction: Identify unintended target interactions
• Personalized Medicine: Predict patient-specific drug responses
• Natural Product Screening: Find candidate drugs from natural compound libraries

This project provides a complete industrial-grade heterogeneous GNN workflow, achieving state-of-the-art performance on the BioSNAP-DTI benchmark and demonstrating the potential of GNNs in the field of drug discovery. For researchers and engineers in computational biology, drug discovery, and GNN fields, this project offers valuable technical references and implementation foundations.