Structure-Preserving Generative Modeling for Spatial Omics
BioFlow keeps histology-conditioned gene-expression trajectories within biologically valid support.
When Biological Data Do Not Live in Euclidean Space
Spatial transcriptomics prediction from histology is commonly formulated with diffusion or flow-matching models. Standard generative trajectories, however, evolve in unconstrained real-valued space and can produce negative gene-expression values that are biologically invalid.
BioFlow
BioFlow incorporates non-negative support directly into the generative dynamics. A reparameterized velocity field preserves valid expression values throughout the probability path instead of repairing invalid outputs afterward.
Why It Matters
The project asks a broader modeling question: if scientific observations live on a constrained space, should those constraints be part of the model’s geometry? Encoding valid support into the generation process improves both scientific consistency and computational efficiency.
Resources
BioFlow was published at MICCAI 2026. The complete citation will be added to the Publications page once its public bibliographic record is available.