DumplingGNN: Hybrid GNN for ADC Drug Discovery

A hybrid graph neural network for antibody-drug conjugate payload activity prediction

Project Overview

DumplingGNN is a novel hybrid Graph Neural Network architecture designed to predict ADC (Antibody-Drug Conjugate) payload activity based on chemical structure. This project addresses the critical challenge of efficient payload activity prediction in ADC drug development.

Key Contributions

Core Algorithm Development

  • Independent Design: Solely designed and implemented the DumplingGNN hybrid architecture
  • Multi-scale Feature Integration: Combined 2D topological and 3D spatial conformation information
  • Architecture Innovation: Integrated Message Passing Neural Networks (MPNN), Graph Attention Networks (GAT), and GraphSAGE

Technical Achievements

  • Model Performance: Achieved 91.48% accuracy, 95.08% sensitivity, and 97.54% specificity on ADC datasets
  • Benchmark Results: Set new SOTA results on MoleculeNet datasets (e.g., BBBP ROC-AUC 96.4%)
  • Interpretability: Implemented attention mechanisms for key substructure visualization

Research Impact

  • Publication: First-author paper published in International Journal of Molecular Sciences (JCR Q1) - View Publication
  • Open Source: Released model code and training data for community use
  • Platform Integration: Successfully deployed algorithm in OmniMedical platform

Technical Innovation

The DumplingGNN architecture represents a breakthrough in molecular property prediction by:

  1. Hierarchical Feature Extraction: Capturing molecular features from local to global scales
  2. Multi-modal Data Integration: Incorporating both 2D and 3D structural information
  3. Explainable AI: Providing interpretable results through attention mechanisms

Awards & Recognition

  • National Scholarship: Individual project recognition
  • University Competition: First Prize in Qiushi Cup Undergraduate Research Competition
  • Algorithm Competition: National Second Prize in Global Campus AI Algorithm Elite Challenge

This is my first full research project, it may be young, but it means a lot to me!

References