A feature matrix stores values describing entities or graph elements. In molecular graph generation, it is reconstructed alongside the adjacency matrix so the generated graph preserves both connectivity and node information. section 5.4.3 section 5.4.4
A feature matrix participates in the generation pipeline as follows:
- Represent each molecular graph with graph data containing node features in `data.x`. section 5.4.4
- Encode graph features into a latent representation using heterogeneous graph-convolution layers and relation-specific edge types. section 5.4.3
- Decode the latent representation into both an adjacency matrix and a feature matrix for each small graph. section 5.4.3
Molecular graph generation
A variational graph autoencoder receives molecular graph data, predicts a feature matrix and adjacency matrix, and trains against the original graph while also predicting QED. The resulting latent space clusters molecules with higher QED, supporting exploration of regions likely to produce candidate molecules. section 5.4.4
