Continuous Representation Of Molecules using Graph Variational Autoencoder
summary
Proposes a graph VAE on 2D adjacency tensor that encodes molecules with stacked GCNs (three layers, first two weights shared), trains on ZINC molecules filtered to <30 atoms (padded to 30), four bond types, and uses a simple inner‑product decoder that reconstructs adjacency only so it cannot yet generate full molecules. A side predictor regularizes the latent space and pools node‑wise latent vectors (node latent dim 16, pooled molecule vector 64) to predict solubility, synthetic accessibility and drug‑likeness; shows better similarity and property prediction than SMILES RNN VAEs but lacks an invertible pooling and full decoder for generation.
AI-written summary · about
chapters (7)
- 0:00Molecular representation
- 4:25Graph convolution basics
- 17:35Variational auto-encoder
- 20:27Decoding and side prediction
- 26:39Model training and architecture
- 30:52Experiments and evaluation
- 34:31Future work and conclusion
related talks
21mGraph Representation Learning with Deep Embedding Approach
57mTutorial on Deep Learning III
01h 20mComparing GNN Approaches for Molecule Identification
23mUsing Graph Nets (GNs) to predict molecular properties
01h 17mDeep generative models for image...
37mFlexibly Fair Representation Learning by Disentanglement
31mML methods to accelerate the drug discovery cycle
01h 23mGraph Neural Networks in Drug Discovery