DHGNN: Dynamic Hypergraph Neural Network for C. elegans Embryogenesis

Caenorhabditis elegans is the only animal whose complete, invariant cell lineage is known, which makes it an ideal system for computational models of development. This project builds a Dynamic Hypergraph Neural Network (DHGNN) that represents the embryo as a time-evolving hypergraph and learns a per-cell embedding across 190 developmental frames. The model integrates three biologically distinct hyperedge types into a single augmented incidence matrix H_aug(t): spatial communities obtained by running DBSCAN on three-dimensional cell positions and rebuilt at every timepoint; lineage parent–daughter bonds derived from the recorded division table and held fixed; and functional feed-forward-loop (FFL) motifs extracted from the adult C. elegans connectome and also held fixed. A Hyper-SAGNN self-attention convolution propagates information from cells to hyperedges and back, and the spatial block is dynamically re-clustered between layers from the learned embeddings using k-means (global) and k-NN (local). The network is trained on a single supervised objective — cell-fate (founder-lineage) classification — using a cell-disjoint train/test split so that accuracy reflects genuine generalisation rather than identity memorisation. On held-out cells the model reaches a fate accuracy of 0.958, and its learned embeddings show strong lineage-coherence (silhouette +0.683 and k-NN purity 0.977, versus −0.151 and 0.721 for a raw-feature baseline), indicating that biologically meaningful structure emerges without being directly supervised.