Dynode
SE(3)-equivariant transformer-based neural ODEs for 3D organogenesis.
Dynode models 3D organogenesis from spatiotemporally resolved single-cell data with an SE(3)-equivariant transformer-based neural ODE. The model captures morphological and expressional changes through optimal transport and supports in silico perturbation for congenital heart disease.
Key ingredients.
- SE(3)-equivariant transformer backbone (rotation- and translation-equivariant feature updates).
- Continuous-time neural ODE for trajectory inference.
- Optimal-transport regularisation tying snapshots through Wasserstein consistency.
- In silico knockouts for hypothesis generation about regulators of cardiac development.
Preprint in preparation; will link here once posted.