LatentFlow Visualizes Molecular GNN Latent Spaces for Chemists

This research introduces LatentFlow, a new visual analytics system developed to help chemists and materials scientists better understand the internal workings of molecular Graph Neural Networks (GNNs). While GNNs are increasingly used to predict molecular properties, comprehending how they internally represent chemical information within their latent spaces (embeddings) is crucial for diagnosing model behavior and ensuring the learned representations align with actual chemical relationships. LatentFlow addresses the limitations of existing analysis methods by providing comprehensive support for exploring latent spaces across different GNN layers and various model states, such as training epochs or configurations. It groups embeddings into clusters and visualizes their evolution using a modified Sankey diagram. Crucially, LatentFlow links these clusters to representative molecules and their shared substructures, allowing domain experts to integrate their chemical knowledge and compare it with the patterns discovered by the model, thereby enhancing model interpretability.