What it is
A multimodal scientific foundation model that reasons directly over protein, molecule, and crystal structures using structure-aware tokens and inspectable traces. Instead of treating structures as text or using opaque task-specific predictors, it tokenizes native 3D/topological structure into evidence units and generates predictions with structure-linked reasoning across proteins, molecules, and crystals.
Key results
- 67 of 86 benchmarks reach state of the art
- Low-homology protein GO CC Fmax improves 0.42→0.55 vs ESM2
- Single-step retrosynthesis accuracy rises 0.63→0.72
Shows that making native structure explicit reasoning evidence improves both accuracy and interpretability across scientific domains.