Model interactions

The Interactions tab of a model's panel lets you set up relationships between the models you have loaded, so that one model's output can depend on another's. Three kinds are available:

• Avoidance — a model can be told to stay away from another within a set distance. For example, a membrane model set to avoid a carbon model within 30 nm will suppress its membrane predictions near the carbon film.
• Colocalization — the opposite: a model's predictions are kept only within a set distance of another's. For example, a membrane-associated protein model colocalizing with a membrane model within 30 nm only predicts near membranes.
• Competition — models can emit and absorb competition. A model that absorbs competition is suppressed to zero wherever an emitting model predicts a higher value for the same voxel, so where two models overlap the more confident one wins.

GUI only

Model interactions exist only in the GUI. There is no command-line equivalent, so they have no effect when you segment a dataset with ais segment.

When to use them

Model interactions were part of the original Ais and are described in the eLife paper; they can be useful for exploring how features relate. Our thinking has since narrowed, though. It helps to separate two jobs: turning a tomogram into a faithful segmentation of a feature, and then using those segmentations — combining features, applying constraints, selecting particular particles. Ais (and easymode) increasingly focus on the first: turn a tomogram into a representation of a specific feature, and do the rest afterwards.

So for most work we would keep the segmentation step clean — segment each feature on its own — and do any constraining or combining in a post-processing script, where it is easier to see and adjust. Play with the interactions if you find them useful, but you generally don't need them to get good segmentations.