🟢 Vault

easymode segment vault

The vault model was trained on manually curated 3D subtomograms that were labelled by a 2D Ais UNet. The Ais net output a shape-based segmentation of vaults; as a result, the 3D easymode network also outputs vault shape, although with strong missing wedge artefacts. This does not really matter for picking.

For initial validation we used dataset EMPIAR-11845, consisting of 152 tomograms of FIB-milled D. discoideum cells. Because this dataset was also a part of the training dataset, there is some contamination between training and validation; we will repeat the validation later with a different dataset. The reason for running validation on this set was that few datasets contain enough vault particles for annotation and averaging.

Vaults remain rare even in this dataset; we found on average between 2 to 3 particles per tomogram, or 348 in total. With D39 symmetry, subtomogram averaging plateaued at a resolution of 13.6 Å. The very cap of the vault, where the D39 symmetry breaks (see Lövestam and Scheres, Structure, 2025), was excluded from the mask.

Example output

Example of easymode segment vault output overlaid on a tomogram from EMPIAR-11899 (FIB-milled D. discoideum), a dataset which was not used to train this model.