Comparing GNN Approaches for Molecule Identification
summary
Tutorial compares GNN pipelines that turn ≈11,630 SMILES into graphs, augment the minority classes to ≈20,800 graphs by node‑index permutations preserving isomorphism, and train GCN, GraphSAGE, GAT and GIN with Optuna tuning on a GPU cluster. Node features come from 150 random walks of 35-step length fed to word2vec producing 64-dim node embeddings combined with one‑hot atom types. GraphSAGE yields the best F1/ROC, but models still miss many minority‑class actives despite augmentation and class weighting.
AI-written summary · about
chapters (10)
- 0:00Introduction
- 2:34Molecule identification motivation
- 7:54Biological background and datasets
- 17:12Data preprocessing and graphs
- 25:34GNN models and training
- 31:04Live demo
- 59:04Evaluation results
- 1:06:20Classification report analysis
- 1:09:12Conclusions and future work
- 1:14:25Q&A
related talks
01h 23mGraph Neural Networks in Drug Discovery
23mUsing Graph Nets (GNs) to predict molecular properties
29mGraph Neural Networks + LLM on
36mAmazing things that happen with Human-AI synergy
38mContinuous Representation Of Molecules using Graph Variational Autoencoder
31mGraph-Based Machine Learning with Python
01h 01mToward Universal Graph Representations: Foundations and Frontiers of Graph Foundation Models
16mPreview: Graph NNs