Lost in the z-axis: Can generative AI reconstruct true 3D tissues from serial sections?

Talk and poster, Genomics at Scale 2026, St John’s College, Cambridge, 23 October 2026

Tissues are 3D, but spatial transcriptomics is measured one 5 µm section at a time. Sequencing every section is slow and expensive, while H&E staining is cheap and routine. In this talk I presented HoloTea, a 3D-aware flow-matching model that predicts spot-level gene expression from H&E, using the matching spots on the sections directly above and below as a guide.

On a human embryo stack of 52 sections (327,532 spots), HoloTea was trained on every other section and predicted the 26 sections in between, recovering anatomical marker genes and tissue domains it never saw. It beats 2D and 3D baselines across three datasets and scales to whole embryos: predicting a section takes 3.7 GB of GPU memory, while the previous 3D method runs out of memory on an 80 GB GPU.

The poster won the People’s Choice Poster Award, voted by the attendees.

Joint work with Mohammad Vali Sanian, Amirhossein Vahidi and Mohammad Lotfollahi.