The training words are steps in one loop. This note defines each one on a 1,000-molecule example, from the batch count and the loss to message passing and a scaffold split.
Oct 24, 2026
Pinning random_state=42 makes a UMAP-based train/test split perfectly reproducible - and that is exactly why it can lull you into a false sense of rigor. Reproducibility guarantees you get the same answer every run; it says nothing about whether that answer is typical. Here's the distinction, why it matters for evaluating GNNs on chemical-domain shifts, and how to fix it.
Mar 28, 2025