ESTGEL · DevoGraph

Cell-fate map of a C. elegans embryo

A spatio-temporal graph network predicts each cell's terminal tissue from its position, size and lineage dynamics — never its name. Play development forward, watch cells divide, and inspect what drives each call.

3D embryo · predicted fate

t = 0 min

drag to rotate · scroll to zoom · click a cell

t = 0 min

What drives the prediction

Attention weights saturate, so importance is measured by perturbation — remove an input and watch the call move.

Do interactions matter? · graph ablation

—

accuracy, full graph

→

—

edges removed

—

of fates flip
without contacts

Removing every cell–cell edge and re-predicting. Confidence drop per tissue → contacts carry real fate signal.

Which features drive each fate · occlusion

Confidence lost when a feature is mean-imputed. Warmer = the feature supports that tissue's call.

Tissue-to-tissue contacts

How often the model sees each tissue in contact with another (log scale). Diagonal = like-fate clustering.

Model performance

val accuracymacro-F1

ESTGEL node classifier · EPIC lineage data fate labels from the invariant C. elegans lineage · trained on held-out embryos