Graph Neural Networks in Drug Discovery
summary
Compares a fingerprint multitask classifier (ECFP, 1,024 bits) to graph approaches on Tox21, 7,800 compounds, 12 tasks, using DeepChem + RDKit; fingerprint model gets test ROC-AUC ~0.67 while an out‑of‑the‑box GCN reaches validation ~0.724 and test ~0.68. Implements a homemade GCN (two graph conv layers, dense 256, batch norm, tanh, batch 32, 10 epochs) and a custom pooling layer testing mean / max / sum pooling; mean slightly outperforms max, sum performs worst. Models are unoptimized but show GCN advantage.
AI-written summary · about
chapters (25)
- 0:00Introduction
- 1:08Why GNNs in drug discovery
- 3:15Molecules as graphs
- 5:29Common GNN tasks
- 7:14GNNs for molecule classification
- 8:17GCNs generalize image convolutions
- 10:03GCN theory and adjacency matrices
- 13:20Toxicity prediction with Tox21
- 15:56Setting up DeepChem for GNNs
- 18:47Handling imbalanced toxicity data
- 20:56Molecular fingerprints and multitask
- 23:04Loading Tox21 dataset
- 26:55Training multitask classifier
- 29:21Out-of-the-box GCN in DeepChem
- 33:55Custom GCN from scratch
- 37:12Graph pooling techniques
- 43:51DeepChem backend and debugging
- 45:57Evaluating custom GCN
- 47:44GNNs in Pharma workflows
- 50:33Mean vs max vs sum pooling
- 56:05Other Graph Neural Networks beyond Graph Convolutional Networks
- 59:46Graph symmetries in molecules
- 1:03:23Summary of Tox21 experiments
- 1:07:21Multitask vs GCN comparison
- 1:10:16Q&A
related talks
01h 20mComparing GNN Approaches for Molecule Identification
31mGraph-Based Machine Learning with Python
23mUsing Graph Nets (GNs) to predict molecular properties
47mPyTorch Geometric for Graph Neural Nets
01h 30mPractical graph neural networks in Python with TensorFlow and Spektral
01h 16mBoot camp on graph foundations models
25mGarrett Goh, Scientist, Pacific Northwest National Lab at MLconf Seattle 2017
16mGraph Network in High Energy Physics