Using Graph Nets (GNs) to predict molecular properties
summary
Graph Nets are applied to molecular graphs and claim to approximate per-atom, per-bond and per-molecule quantum properties with errors within QM uncertainty while delivering an over-500-fold speed up. Implementation rewrites cheminformatics I/O into a TensorFlow/GPU pipeline and evaluates on 350,000 molecules, predicting partial charges with RMSE ≈ 0.02. The work predicts per-bond orders and fragments FDA-like molecules to identify coupled regions, some models are trained and validated, other fragmentation-based models are prepared but not yet fully trained.
AI-written summary · about
chapters (10)
- 0:00Graph Nets for molecular learning
- 1:04Graph Nets structure
- 3:56Graph Nets step-by-step
- 6:02Graph Nets implementation
- 8:44Predicting partial atomic charges
- 13:21Predicting partial bond orders
- 17:21Future applications and summary
- 18:03Q&A
- 19:30Q&A
- 20:43Q&A
related talks
01h 20mComparing GNN Approaches for Molecule Identification
01h 23mGraph Neural Networks in Drug Discovery
31mML methods to accelerate the drug discovery cycle
25mConvergence of AI, Simulations and HPC
36mAmazing things that happen with Human-AI synergy
16mGraph Network in High Energy Physics
01h 01mToward Universal Graph Representations: Foundations and Frontiers of Graph Foundation Models
16mPreview: Graph NNs