Reproduce “LatentFlow: Visual Analytics for Latent Space Analysis in Molecular Graph Neural Networks” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Shiyi Liu, Jiaqing Chen, Nicholas Hadler, Rostyslav Hnatyshyn, Michael W. Mahoney, Talita Perciano, John F. Hartwig, Gunther H. Weber and 1 more
LatentFlow is a visual analytics system designed for computational chemists to understand how molecular information is organized within the latent spaces of Graph Neural Networks. It tracks the evolution of molecular embeddings across different model layers and states, linking these numerical representations to chemical concepts and structures to enhance model interpretability and diagnose behavior.
Key insights
- Molecular GNN latent spaces are highly dynamic, reorganizing from coarse, taxonomy-based structures to property-driven organizations over the course of training, as demonstrated in solubility prediction.
- Data augmentation, such as conformer augmentation, significantly influences latent space organization, leading to improved separation of chemically meaningful groups related to target properties.
- LatentFlow helps connect abstract embeddings to specific chemical features (e.g., substructures, molecular size, halogenation) that drive the observed organization of molecules, providing actionable chemical interpretability.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.